AI for Bioscience — The Market Radar 174 companies across 20 segments, every row carrying its sourced traction and a citation. Each segment is a layer — open one for where the gaps are. Sources and what this is not
August 2026 vintage 174 of 174 sourced
98
gaps identified— show all
The shape of the market Estimated segment size (USD)
Market size Companies Gaps
What this shows: how big each segment is, per the best published estimate we could source. Therapeutics — preclinical is the largest at $8.2B and 94x the smallest sized segment (Pathology & spatial biology). Scopes differ, so treat each bar as its own estimate.
Therapeutics — preclinical $8.2B 26
Lab systems of record $6.3B 25
Autonomous labs $6.0B 25
Industrial biology $5.0B 25
Scientific data infrastructure $4.5B 26
Target discovery $4.1B 25
Trial design & digital twins $4.0B 24
Proprietary data assets $3.2B 26
Biomanufacturing operations $2.4B 25
Cell-therapy manufacturing $2.1B 26
Foundries & training data $2.0B 26
Wet-lab validation $1.9B 25
Computational chemistry $1.6B 25
Protein design software $1.5B 25
Programmable medicine $470M 26
Omics & single-cell models $379.95M 25
Pathology & spatial biology $87.2M 24 Bar colour = how crowded the segment is saturated contested open
Bar height is compressed (square-root) because the sourced range spans roughly 94× — on a straight scale most segments would be invisible. Read the printed figure, not the height. 3 of 20 segments carry no bar: no credible published figure exists, and none was invented. Per-segment market figures are third-party estimates, each sized on its own scope and vintage — they are NOT a comparable series and must never be summed. Every figure states what it actually measures and how that differs from the segment. Three of the 20 segments carry no figure: no credible published estimate matches them, and none was invented or borrowed from a broader market.
Who is building what Every company in the corpus, stacked by segment
174 of 174 companies in scope · 20 of 20 segments shown · layers ordered and sized by company count — the corpus carries no market caps, so a box earns attention by its evidence: open one for the sourced traction.
S1 Foundation models contested 7 5 gaps identified — open for the gaps, incumbents and entry risks. Chai Discovery US · Builds frontier biomolecular structure and design models (Chai-1, Chai-2) and licenses them to drugmakers rather than running its own pipeline. EvolutionaryScale US · Built ESM3, a generative protein language model spanning sequence, structure and function, released with an open-weights research tier and an API; distributed to researchers and pharma rather than developing its own drugs. Profluent Bio Other · Trains protein language models that design working gene editors, antibodies, antigens and enzymes; releases flagship designs openly and builds on them commercially. Basecamp Research UK · Collects biodiversity samples under benefit-sharing agreements to build a proprietary protein-sequence database, then trains structure-prediction and enzyme/gene-editor design models on it. BioMap Other · Builds xTrimo, a cross-modal life-science foundation model family covering protein, DNA, RNA, cell and text modalities, and sells model access plus co-development to pharma. Helixon Other · Deep-learning platform for protein structure prediction and antibody sequence design, coupled to high-throughput wet-lab validation; monetizes by out-licensing designed biologics. Deep Genomics Other · Trains RNA-biology foundation models — BigRNA plus DeepADAR for RNA editing and REPRESS for microRNA-driven degradation — that predict tissue-specific RNA behaviour from sequence. + Gap Nobody in this segment sells independent third-party evaluation of biomolecular models The rows supply the evidence themselves: Chai-2's 16% hit rate is a company preprint the row calls not independently replicated, Basecamp's 6x/10x multiples are explicitly "company benchmarks, not third-party evaluations", and BioMap's parameter count is a vendor claim — a buyer comparing them has no neutral scorer. Caveat: this is an absence from this radar's snapshot only; the closest owners on the radar sit in "Closed-loop wet-lab validation as a service" (Adaptyv Bio, LabGenius, Arctoris, Tierra Biosciences, Nuclera), who may already run de facto head-to-head benchmarking under contract, and academic efforts like community structure-prediction assessments would never appear in a commercial radar at all. + Gap No row here is DNA- or regulatory-genome-first, and RNA has a single owner Deep Genomics is the only RNA-native model company in the segment (BigRNA, DeepADAR, REPRESS) and is also the one row with no verifiable financing, so the modality with the thinnest coverage is also the least capitalised. Caveat: thin in THIS segment, not absent from the radar — "Multi-omics & single-cell foundation models" (Immunai, Bioptimus, Synthesize Bio, CytoReason) and "Programmable-medicine design platforms" (Moderna, Inceptive, Dyno Therapeutics, Form Bio) both carry nucleic-acid modelling work in this same snapshot. + Gap Nobody in this segment has shown a priced, metered self-serve model business Every disclosed monetization route is something else: pharma co-development (BioMap's $10M Sanofi upfront, the rest milestone-contingent), asset out-licensing (Helixon), or free release (Profluent's OpenCRISPR-1, ESM3's open weights on AWS and NVIDIA BioNeMo) — and no row states product revenue. Caveat: absence of disclosure is not absence of revenue, and the radar does place the commercial-software layer elsewhere, in "Protein & enzyme design software for enterprise R&D" (Cradle, Latent Labs, Nabla Bio and seven others), so the seat-and-subscription business exists in this snapshot — just not in this segment. + Gap Only one row owns its own wet lab Helixon is alone in coupling the model to high-throughput experimental validation; the other six ship weights, papers or licences and depend on someone else to run the experiment that would confirm the claim — which is the same gap that makes the vendor benchmarks uncheckable. Caveat: the capability is emphatically owned on this radar, just outside this segment — "Autonomous labs & self-driving experimentation" (Lila Sciences, Automata, Atinary Technologies, Tetsuwan Scientific) and "Cell-programming foundries & lab-generated training data" (Ginkgo Bioworks, Tahoe Therapeutics, Somite AI, Asimov, Bota Biosciences) — so the white space is vertical integration inside a model vendor, not the existence of the service. + Gap No one sells provenance, benefit-sharing or licensing compliance for biological training data Basecamp Research is the only row whose sourcing model is legal as much as technical — biodiversity samples collected under benefit-sharing agreements — and it keeps that apparatus for itself rather than offering it as infrastructure to the other six, whose training-corpus provenance is simply not described in any row. Caveat: an absence from this radar's snapshot, and the adjacent "Proprietary biological data assets for model training" segment (Variant Bio, deCODE genetics, MedGenome, Nightingale Health, A-Alpha Bio, Protai) holds companies that own consented data — owning compliant data is not the same as selling the compliance layer, but a radar organised by company rather than by function would not surface such a layer even if it existed. S2 Proprietary data assets open 6 5 gaps identified — open for the gaps, incumbents and entry risks. A-Alpha Bio US · Runs AlphaSeq, a yeast-based assay that measures millions of quantitative protein-protein binding affinities, then trains AlphaBind models on that proprietary interaction data. Variant Bio US · Assembles consented genomic and phenotypic data from under-studied populations with outlier traits, and mines it for drug targets under a benefit-sharing agreement with each community. Nightingale Health EU · Runs an NMR blood-analysis platform that produces standardised metabolic biomarker profiles at biobank scale, then licenses the resulting dataset and derived risk scores. MedGenome Asia · Genomic diagnostics and research company whose durable asset is the largest variant database for South Asian populations, sold to pharma for target discovery and validation. Protai Asia · Generates its own large-scale mass-spectrometry proteomics data from clinical tissue and trains models on it to find protein-level drug targets that genomics misses. deCODE genetics Other · Holds and analyses population-scale genotype and phenotype data for a large share of Iceland's population — the archetype of a proprietary human-genetics training asset. + Gap No row is a cross-owner aggregator — each of the six owns exactly one asset and sells it bilaterally itself, so nobody assembles several owners' assets into one licensable training corpus. Checked against the corpus index: the federated-access plumbing does exist elsewhere in this radar under Scientific data infrastructure (Lifebit, DNAnexus, Velsera, Elucidata), so the gap is a data OWNER playing aggregator, not the technology. Caveat: this is one August 2026 snapshot of 174 companies, and consortium efforts like GenomeAsia 100K (which MedGenome co-founded) are a non-commercial version of the same idea. + Gap Nobody here sells treatment-response data — molecular profile linked to what happened when a specific drug was given All six assets are observational or in-vitro: population genetics (Variant Bio, MedGenome, deCODE), baseline biomarker profiles (Nightingale), binding measurements (A-Alpha), tumour tissue proteomics (Protai), and every named pipeline is preclinical, so no row has dosed-patient data of its own. Caveat: this radar covers the adjacent job in a different segment — Sponsor-side trial design & digital twins (Medidata AI, Unlearn.AI, QuantHealth, Pathos AI, Genialis) works with sponsor trial data — so the absence is of a licensable trained-on asset, not of anyone touching trial data. + Gap Nothing in this segment attests to data quality independently The strongest claims in the rows are self-reported (Nightingale's 2.8M+ samples and 900+ publications are marked vendor figures; MedGenome's 'largest' is unaudited), and the only externally-anchored counts come from a customer, not an auditor — A-Alpha's affinity volumes via Department of Defense contracts, Nightingale's 500,000 profiles via UK Biobank. No row in this radar's 174 companies benchmarks or certifies someone else's biological dataset. Caveat: this may be an in-house buyer function or a standards-body job rather than a venture-scale company, and the radar only shows what got funded. + Gap Variant Bio built consent and benefit-sharing machinery across nine countries and territories including Maori partners, and nobody sells that governance layer to anyone else — it exists once, inside one company, as a contract term rather than a product. No other row in this segment or elsewhere in the radar offers sovereignty-compliant consent, provenance and revenue-sharing infrastructure to data owners. Caveat: absent from THIS snapshot only; this work is often done by law firms, ethics boards and national biobank authorities that a company radar would never list. + Gap Every asset here is human clinical or in-vitro; no row holds environmental, microbial or broader non-human biodiversity data. Checked against the corpus index before claiming absence: this radar does contain exactly that business — Basecamp Research sits in the Biomolecular foundation models segment — so the finding is that non-human data ownership is classified as a model company rather than a data-asset company in this snapshot, which itself says the pure data-licensing model has not been tried there. Caveat: segment boundaries are the corpus's own judgement, and Ginkgo Bioworks and Tahoe Therapeutics in the Foundries & lab-generated training data segment generate non-human data too. S3 Protein design software contested 10 5 gaps identified — open for the gaps, incumbents and entry risks. Cradle EU · Sells protein-engineering software: customers upload a sequence, the models propose variants, and an in-house wet lab closes the loop on measured data. InstaDeep UK · Sells DeepChain, a generative protein and nucleotide design platform (ProtBFN, AbBFN, Nucleotide Transformer), now opened to external partners alongside its parent's internal use. Latent Labs UK · Sells Latent-X, a browser-based generative model that designs novel protein binders with atomic structures, aimed at biotechs and pharma without in-house design teams. Nabla Bio US · Sells de novo antibody and protein design through its JAM generative platform, run as paid discovery partnerships rather than owned assets. Scala Biodesign Asia · Sells ScalaOS, computational protein-redesign software that stabilises and optimises existing proteins for manufacturability and expression. BigHat Biosciences US · Runs the Milliner platform — ML models coupled to a high-speed synthetic-biology wet lab — to design antibodies for pharma partners on multiple properties at once. Galux Asia · Sells GaluxDesign, a de novo antibody design platform, to pharma partners as co-development deals. Enzymit Asia · Designs novel enzymes computationally and sells them as biocatalysts for cell-free bioproduction across pharma ingredients and food. Biomatter Other · Sells generative-AI enzyme design (Intelligent Architecture platform) as a service — customers specify a reaction, Biomatter designs the enzyme. Peptone UK · Runs Oppenheimer, a protein-modelling platform combining biophysics experiments, molecular dynamics and ML to make intrinsically disordered proteins tractable as drug targets. + Gap Nobody in this segment sells or publishes an independent, audited head-to-head benchmark of design-model performance The evidence: every performance figure in these ten rows is the vendor's own -- Cradle's 'up to 12x / up to 90% lower cost' with no published methodology, Scala's nine-of-top-20 with no customer named, InstaDeep's ~10,000 designs per day from BioNTech's own AI Day -- so a buyer has no way to rank them. Caveat: this is an absence of an arbiter, not of measurement capacity, and only within this radar's snapshot -- the radar's Closed-loop wet-lab validation as a service segment (Adaptyv Bio, LabGenius, Arctoris, Tierra Biosciences) already supplies the wet-lab measurement such a benchmark would run on, and an academic or consortium benchmark could exist outside this corpus entirely. + Gap No row here designs a non-protein modality -- nucleic acids, mRNA payloads, cell or gene constructs The evidence: InstaDeep's Nucleotide Transformer is the only nucleotide capability named, and it sits inside BioNTech rather than being sold as enterprise design software. Caveat: this is a boundary of this segment, not a gap in this radar -- the radar covers it in Programmable-medicine design platforms (Moderna, Inceptive, Dyno Therapeutics, Form Bio), so a new entrant would be crossing into an occupied segment, not an empty market. + Gap No row differentiates on a disclosed proprietary measurement corpus -- including the failed designs -- as the asset rather than the model. The evidence: Cradle and BigHat own wet labs but describe them as loop-closing infrastructure, and no row in this segment discloses a data asset, its scale, or exclusive rights to it. Caveat: within this radar the data-asset business is deliberately filed elsewhere -- Proprietary biological data assets (A-Alpha Bio, Basecamp Research) and Cell-programming foundries & lab-generated training data (Ginkgo Bioworks, Tahoe Therapeutics) -- so the absence is of a design vendor that sells the data moat, in this snapshot only. + Gap No row shows a priced, self-serve, seat-based software motion with disclosed revenue The evidence: Latent Labs shipped a browser platform in July 2025 aimed at teams without in-house design capability but discloses no pharma revenue, and every other row monetises through enterprise deals, co-development (Galux with Celltrion and LG Chem) or partnerships (BigHat with Eli Lilly). Caveat: absence of disclosure is not proof of absence of the motion, and this radar puts the seat-priced lab-software business in Lab systems of record (Benchling, Dotmatics) -- adjacent buyers, different product. + Gap No China-based protein or enzyme design vendor appears in this segment; Galux (Seoul) is the only row outside the US, Western Europe and Israel. The evidence: the ten HQs are Amsterdam/Zurich, London x3, Cambridge MA, Tel Aviv, San Mateo, Seoul, Rehovot and Vilnius. Caveat: this is a gap in this segment of this radar, not in the market -- the radar itself lists China-based players one segment over (BioMap and Helixon under foundation models, XtalPi and DP Technology under computational chemistry, Bluepha under industrial biology), so the capability is present in the corpus and simply not competing for this segment's enterprise-design budget in this snapshot. S4 Computational chemistry saturated 14 5 gaps identified — open for the gaps, incumbents and entry risks. Schrodinger US · Sells physics-based molecular simulation software to nearly every large drugmaker, and co-develops or self-develops candidates using the same stack. DP Technology Other · Sells Hermite, a computer-aided drug design platform (Uni-FEP, Uni-Mol, Uni-EM), alongside the Bohrium research cloud and Uni-Lab lab OS. Cresset UK · Sells a licensed desktop/cloud molecular design suite — Flare (docking, MD, FEP), Spark (scaffold hopping), Blaze (virtual screening) — plus a discovery-services CRO arm. Rowan Scientific US · Sells a cloud computational-chemistry platform that pairs quantum-chemistry calculations with machine-learned interatomic potentials (pKa, conformers, binding affinity) on a self-serve basis. Iktos EU · Licenses Makya, generative de novo small-molecule design software for multi-parameter optimisation, as SaaS, on-premise or in a customer VPC. Qubit Pharmaceuticals EU · Sells Atlas, a molecular-simulation and drug-design software suite built for supercomputers and hybrid quantum/HPC hardware. OpenEye, Cadence Molecular Sciences US · Sells the Orion cloud molecular design platform plus small-molecule and antibody discovery suites and the OEChem toolkits, licensed to pharma R&D organisations. Optibrium UK · Sells StarDrop, subscription small-molecule design and multi-parameter-optimisation software, with deep-learning imputation modules layered on top. PostEra US · Sells Proton, generative chemistry with synthesis-aware design, to pharma as multi-year AI discovery collaborations, with an option for partners to license the tech in-house. SandboxAQ US · Sells AQBioSim — Large Quantitative Models for molecular simulation, including an absolute-FEP binding-affinity engine and physics-refined co-folding — to biopharma R&D teams. Achira US · Builds foundation simulation models — geometric deep learning fused with quantum chemistry and statistical mechanics — sold as molecular 'world models' for drug and materials design. Deep Origin US · Sells a drug-discovery platform pairing multi-scale molecular simulation (docking, MD) with ML prediction, plus ComputeBench and the Balto assistant, to pharma programs. XtalPi Other · Sells quantum-physics plus AI drug and materials discovery as software and services, including a retrosynthesis system and structure prediction, to pharma and industrial clients. Elix Asia · Sells Elix Discovery, an AI drug-discovery platform whose models are trained by federated learning across pharma partners' confidential compound data, kept behind each firewall. + Gap Nobody in this segment sells an independent prospective benchmark of binding-affinity accuracy Every performance number in these 14 rows is self-reported — Qubit's 2x faster candidate selection and 10x cheaper, SandboxAQ's 4x acceleration, Cresset's continuous FEP development since 2019, DP Technology's 50 pipeline projects — and each row says so. No company anywhere in this radar's 20 segments sells third-party evaluation of scientific prediction accuracy. Caveat: this is an absence in this snapshot of 174 companies, not in the world — academic FEP benchmarking consortia and blind challenges exist outside the radar's company scope, and a benchmark business may be a service line rather than a fundable company. + Gap No row couples a physics platform to experimental verification it owns The prediction is sold here; the wet-lab loop that would falsify it sits in OTHER segments of this radar — Adaptyv Bio, LabGenius, Arctoris, Tierra Biosciences and Nuclera in closed-loop wet-lab validation, and Lila Sciences, Automata, Atinary and Tetsuwan Scientific in autonomous labs. So the gap is the seam, not the capability: nobody sells simulation with a contractually-owned confirmation cycle attached. Caveat: PostEra, SandboxAQ and Deep Origin run partner programs where a partner's wet lab closes the loop, so the function exists in these deals without being a product anyone here sells. + Gap No row sells a simulation package built to be accepted by a regulator Deep Origin's $30M ARPA-H CATALYST award is the only regulatory-adjacent item in the segment and the row describes it as building in-silico models that could replace animal testing — a research grant, not a qualified submission path, and its DLBCL work is explicitly a rationalisation rather than a designed candidate. Regulatory-grade in-silico work in this radar lives in the sponsor-side trial design segment (Simulations Plus, Unlearn.AI, QuantHealth, VeriSIM Life), not here. Caveat: the absence is of a molecular-design platform carrying a named regulatory acceptance route in this snapshot; the pharma-simulation regulatory literature and any vendor's private qualification work are out of the radar's view. + Gap Every agentic assistant in this segment is welded to its own vendor's stack, and nobody sells a cross-vendor one Schrodinger launched Bunsen with a Bristol Myers Squibb early-access agreement, Deep Origin ships Balto, and SandboxAQ made AQBioSim callable from Claude in May 2026 — three assistants, three walled engines, no row offering an orchestration layer that drives a customer's mixed estate of Schrodinger, Cresset and OpenEye licences. Caveat: absence in this radar's snapshot only, and it is genuinely unclear whether incumbents would tolerate such a layer — Optibrium's 160-organisation base and Pfizer's enterprise licence are exactly the accounts a neutral layer would need, and they are the incumbents' to gate. + Gap Only one row is priced for a scientist with a credit card Rowan Scientific is the sole self-serve entry, at $2.1M pre-seed with self-serve users and no disclosed enterprise contracts; the rest are enterprise licences (OpenEye/Pfizer, Optibrium at ~$10k/year and up), on-prem or VPC deployments (Iktos's Makya), or multi-year discovery collaborations (PostEra, SandboxAQ, XtalPi). Usage-priced simulation as an API sits unoccupied at scale in this snapshot. Caveat: Rowan already occupies the shape at pre-seed scale and shipped FEP as a product feature in March 2026, so this is a claim about the absence of a scaled owner, not an empty category — and Iktos's AWS Marketplace listing shows incumbents can reach for the same channel. S5 Therapeutics — in the clinic contested 17 5 gaps identified — open for the gaps, incumbents and entry risks. Recursion Pharmaceuticals Other · Runs automated cell-imaging experiments at industrial scale and mines the resulting 'phenomics' maps to pick drug targets and molecules, then develops the resulting candidates itself. Insilico Medicine US · Generative models pick the target and design the molecule (Pharma.AI); develops its own pipeline and licenses the software to other drugmakers. Iambic Therapeutics US · Neural structure-prediction and property models drive a closed-loop design-make-test cycle for its own oncology small molecules. Generate Biomedicines Other · Generative protein models design antibodies and other protein therapeutics from scratch; runs them through its own clinical development. Absci Other · Zero-shot generative antibody design paired with in-house wet-lab validation; develops its own candidates and partners the platform. Enveda Biosciences Other · Uses machine learning on mass-spectrometry data to work out the chemical structures inside plant and microbial extracts, then develops the resulting natural-product molecules as drugs. Relay Therapeutics US · Designs small molecules by simulating how target proteins move (its Dynamo platform blends molecular-dynamics simulation with machine learning), then develops them itself in oncology. Verge Genomics US · Finds neurodegeneration drug targets by running machine learning over its own human post-mortem brain tissue multi-omics, then develops the resulting small molecules. Lantern Pharma US · Runs its RADR machine-learning platform over genomic and drug-response data to pick which patients a rescued or in-licensed oncology molecule will work in, then trials it in that biomarker-selected group. Nimbus Therapeutics US · Designs small molecules against protein targets others consider undruggable, using physics-based computational chemistry (built on Schrodinger's stack) rather than screening. Eikon Therapeutics US · Tracks single protein molecules inside living cells with super-resolution microscopy at industrial scale and applies machine learning to the resulting movies to find drug targets and leads; runs its own oncology and neuro clinical programmes. Frontier Medicines US · Maps reactive pockets across the proteome with chemoproteomics and uses machine learning over that data to design covalent small molecules against targets with no obvious binding site. Regor Therapeutics Other · Runs an in-house computation-plus-structure engine (rCARD) to prototype and select small molecules fast, developing its own metabolic and oncology candidates. MindRank Other · Designs small molecules with its own generative models and develops them through the clinic itself, led by an oral GLP-1 agonist for obesity and type 2 diabetes. METiS TechBio (METiS Pharmaceuticals) Other · Uses AI to design drug formulations and delivery systems (its AiTEM platform), reformulating and improving molecules into its own clinical assets. Pharos iBio Asia · Discovers small-molecule oncology drugs with its Chemiverse platform, which screens and filters candidates on predicted absorption and toxicity, and develops them itself. Valo Health US · Runs an end-to-end computational platform (Opal) over human clinical and molecular data to pick targets, biomarkers and trial populations for its own cardiometabolic and ophthalmology assets. + Gap Nobody in this segment runs a comparison that isolates whether AI origin changed the outcome — no row reports a head-to-head against a conventionally-discovered comparator, and only Insilico's rentosertib has a peer-reviewed placebo-controlled readout at all. The nearest thing in this radar sits in another segment: Intelligencia AI and Unlearn.AI under 'Sponsor-side trial design, digital twins & translational AI' predict trial success for sponsors, which is forecasting, not an audit of AI-origin attribution. Caveat: this is an absence in this radar's August 2026 snapshot of published evidence, not proof that no company holds internal comparator data it has chosen not to publish. + Gap No dosed AI-designed asset in this segment is anything other than a small molecule, an antibody, or a reformulation — no cell, gene, or mRNA therapeutic appears here. This radar does carry those design platforms, but in a separate segment ('Programmable-medicine design platforms (mRNA, cell & gene)': Moderna, Inceptive, Dyno Therapeutics, Form Bio), and none of them shows up in this clinical-stage list. So the gap is specifically an AI-designed programmable medicine that has reached a dosed patient under an AI-native developer's own banner. Caveat: Moderna plainly runs clinical programmes; the radar simply does not place a dosed AI-designed programmable medicine in this segment. + Gap Zero rows work in infectious disease, antimicrobial resistance, psychiatry, or vaccines The 17 assets cluster into oncology (Recursion, Iambic, Relay, Lantern, Nimbus, Eikon, Frontier, Pharos iBio), metabolic/obesity (Regor, MindRank), immunology and dermatology (Absci, Enveda, Generate), and a thin neuro tail (Verge in ALS, METiS in pseudobulbar affect). Caveat: no other segment of this radar covers anti-infectives either, so the absence is radar-wide — but the reimbursement economics of AMR are a well-known cause, and a 174-row snapshot is not a market census. + Gap No row offers clinical development as a service to AI-first discovery shops — every company here that discovers also pays to develop, and the price is explicit: Generate earmarked roughly $300M of its $400M IPO for Phase 3 alone, and Eikon earmarked ~$100M for a single lead. Nothing in this segment or elsewhere in the radar is structured as 'bring us your AI-designed asset and we run the trials'. Caveat: big pharma performs exactly this function through licensing (Genentech optioning a Recursion target, Takeda with Iambic, Lilly with Nimbus, Amgen and Novartis with Generate) — so it is a deal structure that already exists, not a missing capability, and only the standalone company form is absent from this radar. + Gap Nobody publishes or pools the failure data, even though this segment generates plenty of it — Recursion's four halted 2026 programmes, Valo's withdrawn OPL-0301 Phase 2, and Verge's unconfirmed readout are each visible only as a line in a press release or an inference from silence. There is no negative-results resource for AI-originated clinical assets in this segment, and the radar's nearest neighbour ('Proprietary biological data assets for model training': A-Alpha Bio, Variant Bio, deCODE genetics) sells training data, not clinical attrition data. Caveat: absence in this radar is not evidence that no consortium or regulator-held dataset exists off-radar, and the commercial incentive against publishing one's own failures is obvious. S6 Therapeutics — preclinical saturated 9 5 gaps identified — open for the gaps, incumbents and entry risks. Isomorphic Labs UK · Alphabet's drug-design company, built on DeepMind's AlphaFold structure-prediction lineage; designs small molecules for its own oncology and immunology pipeline and for pharma partners. Xaira Therapeutics US · Vertically integrated AI-first biotech: builds de novo binder-design and virtual-cell models alongside its own wet lab, aiming at targets conventional antibody discovery cannot reach. Terray Therapeutics Other · Runs miniaturised chemistry assays that generate billions of measured molecule-target interactions, then trains chemistry foundation models (COATI) on its own data to design small molecules. Ochre Bio UK · Deep-phenotypes donated human livers on perfusion machines, applies machine learning to the resulting genomics, and develops RNA therapies for advanced liver disease from the targets it finds. Aqemia EU · Generates small molecules atom by atom using statistical-physics-based affinity prediction instead of large experimental training sets, running an in-house oncology pipeline alongside pharma deals. Genesis Molecular AI (Genesis Therapeutics) US · Generates small molecules with its GEMS deep-learning models against hard oncology and immunology targets, running both an in-house pipeline and pharma collaborations. insitro US · Trains machine-learning models on cell-imaging and human genetic data to pick drug targets and patient segments, then develops medicines against them with partners and in-house. CHARM Therapeutics UK · Designs small molecules using DragonFold, its deep-learning system that predicts the 3D structure of a protein and a candidate ligand together, and develops the results itself in oncology. Superluminal Medicines US · Generates ensembles of shapes a G-protein-coupled receptor can adopt and uses generative AI plus giga-scale virtual screening against them to design oral small molecules for obesity and cardiometabolic disease. + Gap Nobody in this segment validates anyone else's preclinical package CHARM's resistance-mutation potency and animal tumour-regression results are marked company-reported in its own row, Superluminal's and Ochre's headline deal values are milestone-loaded, and no row offers or buys independent verification of another's AI-derived candidate. Caveat: the absence is partly structural — all nine rows are drug developers, not auditors — and the nearest job on this radar sits in a different segment, 'Closed-loop wet-lab validation as a service' (Adaptyv Bio, LabGenius, Arctoris, Tierra Biosciences, Nuclera), whose rows I was not given; this is a claim about these nine rows in this snapshot, not about the market. + Gap No row sells, publishes or is paid for evidence that the AI route beat the conventional one Every traction field records rounds, partnerships and milestone payments — Terray's BMS discovery milestone, insitro's $25M from BMS, Genesis's $150M in Incyte upfronts including $40M of equity — and none records a head-to-head against a conventionally-designed comparator or a pre-registered success criterion. Caveat: traction fields capture what gets announced, so a company could be running such a comparison unpublished; the honest reading is that this snapshot's segment does not market the proof, not that nobody generates it. + Gap Nobody owns the IND-to-first-dose execution layer that this segment visibly needs Isomorphic slipped its first trial by a year without naming an indication, Aqemia's late-2025/early-2026 guidance passed with no dosing confirmed, and Terray, CHARM (Q1 2026) and Superluminal sit in the same position at the August 2026 check — five slipped or unconfirmed clinical entries, and not one row whose product is getting AI-derived molecules through that gate for others. Caveat: this radar carries a separate 'Sponsor-side trial design, digital twins & translational AI' segment (Turbine, Unlearn.AI, QuantHealth, Medidata AI and seven others), so clinical-side capability is on the radar elsewhere and I have only names for it; what is absent is an owner of the operational hand-off inside these nine rows. + Gap Indication coverage is narrow and self-similar oncology at Isomorphic, Aqemia, Genesis and CHARM, cardiometabolic and obesity at Superluminal and in insitro's BAT-01, liver at Ochre, immunology and inflammation at Xaira, ALS through insitro's BMS deal. No row here runs an anti-infective, antimicrobial-resistance, psychiatric or rare-disease programme — areas where targets are hard to find but trials are comparatively cheap to run, which is exactly the shape a capital-constrained entrant would want. Caveat: this is an absence in these nine preclinical rows only; the radar's 17-name clinical-stage therapeutics segment and its 12-name target-discovery segment may cover those indications, and I was given only their company names. + Gap Only one row treats the data it generates as a product Ochre Bio licences its liver phenotyping data to GSK for up to $37.5M, while Terray's billions of measured molecule-target interactions, Xaira's in-house wet lab and insitro's cell-imaging and human-genetic data all stay captive to their owners' own models. Caveat: this radar puts data-as-a-business in its own segment, 'Proprietary biological data assets for model training' (A-Alpha Bio, Variant Bio, Nightingale Health, MedGenome, Protai, deCODE genetics), so the job is covered on the radar — the observation is only that inside this segment one company of nine monetises it, and an entrant doing so would be stepping into that other segment's competition. S7 Programmable medicine open 4 5 gaps identified — open for the gaps, incumbents and entry risks. Moderna US · Designs and manufactures mRNA medicines on a digitised platform, using sequence-design algorithms plus a company-wide generative-AI layer across research, manufacturing and commercial functions. Inceptive US · Designs RNA molecules — 'biological software' — with transformer models, sold as foundation-model collaborations to RNA therapeutics developers. Dyno Therapeutics US · Designs AAV gene-therapy capsids with ML (LEAP) trained on billions of in-vivo sequence-function measurements, and licenses the resulting vectors to gene-therapy developers. Form Bio US · Sells FormSightAI and FormManufacturing — AI screening of AAV vector designs for yield and contaminant reduction, plus codon optimisation, capsid analysis and batch QC for cell and gene therapy developers. + Gap The 'cell' in this segment's own title has no owner All four rows design nucleic-acid payloads or the vector that carries them - Moderna and Inceptive on RNA, Dyno and Form Bio on AAV - and none designs the cell itself: CAR or TCR construct choice, iPSC differentiation programme, donor selection. Caveat: this is an absence in this radar's August 2026 snapshot, not in the world, and neighbouring segments hold adjacent pieces - cell-therapy manufacturing automation is its own segment (Cellares, Cellino, Ori Biotech, Cellular Origins, Multiply Labs, CellFE) and cell-programming foundries another (Ginkgo Bioworks, Somite AI, Asimov). What is missing here is the AI design layer for a cell therapy, not tooling around one. + Gap Nobody in this segment sells non-viral delivery design - ionizable lipid and LNP formulation as a designed artefact Moderna ships LNP-delivered mRNA at commercial scale but is described by its sequence-design platform and a company-wide generative-AI layer, not a delivery-design product; Dyno and Form Bio both work on viral capsids, a different delivery problem. Caveat: the computational chemistry segment of this same radar holds fourteen physics-based design houses (Schrodinger, XtalPi, Iktos and others) that could in principle reach lipid chemistry, and a 174-row snapshot may simply not have picked up a specialist - the claim is about this radar, not the market. + Gap No row here designs a gene editor or its guide - base and prime editor protein design, guide selection, off-target prediction - even though editing is the third programmable modality alongside mRNA and AAV. Dyno designs the vehicle rather than the cargo, and Inceptive's described work is RNA therapeutics modelling, not editing. Caveat: adjacent segments of this radar hold sequence and protein foundation-model companies (Profluent Bio, EvolutionaryScale, Deep Genomics under foundation models; Nabla Bio, Cradle under protein design) any of which could already reach into editor design - so the honest statement is that no dedicated editor-design platform appears in THIS segment of THIS snapshot. + Gap There is no arm's-length validation of designed-vector or designed-RNA performance available to a buyer in this segment Every efficacy number in the rows is self-reported: Form Bio's 4.4x yield improvement is a vendor claim, its revenue is a Tracxn estimate rather than company-reported, and Moderna's adoption figures are joint claims with its own AI vendor. A developer comparing two capsid sources has no neutral bench. Caveat: closed-loop wet-lab validation as a service exists as its own segment in this radar (Adaptyv Bio, LabGenius, Arctoris, Tierra Biosciences, Nuclera), so what is absent is an AAV/RNA equivalent inside this snapshot - the job exists, just not for this modality here. + Gap Nobody carries a design through to the regulatory evidence a gene-therapy IND actually needs - potency assay strategy, comparability across process changes, immunogenicity and the redosing argument - as a product. Form Bio reaches furthest with yield, contaminant reduction and batch QC, and stops at manufacturing science; Dyno hands over a licensed capsid and its partner owns the filing. Caveat: GxP-facing work sits in other segments of this radar (AI for GxP biomanufacturing operations - Aizon, Apprentice.io, Basetwo; Ketryx under scientific data infrastructure), so this reads as an unbridged seam between the design rows here and the compliance rows there in this snapshot, not as proof that no such vendor exists. S8 Target discovery contested 12 5 gaps identified — open for the gaps, incumbents and entry risks. Causaly UK · Reads the biomedical literature and databases at scale into a causal knowledge graph, then lets researchers query it with AI agents to find and de-risk drug targets, biomarkers and competitive pipeline intelligence. BenevolentAI UK · Mines literature, omics and clinical data into a biomedical knowledge graph to generate and rank drug-target hypotheses, sold both as a platform and worked through its own pipeline. Relation Therapeutics UK · Runs a lab-in-the-loop loop — patient-tissue single-cell data, perturbation experiments and the MORGAN foundation model — to find and validate targets. Healx UK · Builds a rare-disease knowledge graph and uses it to pick drug/combination candidates, then develops them in-house. Standigm Asia · Knowledge-graph target identification (Standigm ASK) chained to generative compound design and repositioning. BenchSci Other · Reads the biomedical literature and figure images with ML and returns experiment-ready target, model and reagent evidence to preclinical scientists. Biorelate UK · Curation engine (Galactic AI) that mines causal biomedical relationships out of papers and patents into a knowledge graph pharma teams query for targets. ONTOFORCE EU · Sells DISQOVER, an ontology-based knowledge graph that fuses ~70 public life-science sources with a pharma's internal data for target and indication search. FRONTEO Asia · Applies its KIBIT NLP engine (Cascade Eye) to unstructured literature to generate novel target and indication hypotheses for pharma partners. Genomenon US · Indexes the full-text genomics literature into gene/variant evidence sets (Mastermind) that biopharma uses for target rationale, patient-population sizing and regulatory dossiers. nference US · Turns de-identified health-system records and literature into target, biomarker and indication evidence for life-science customers (nSights). Excelra Asia · Curates GOSTAR, a structure-activity/ADMET database extracted from patents and papers, and licenses it as the training and lookup layer under other people's discovery AI. + Gap Nobody in this segment sells an audited prospectively-scored hit rate for the targets its graph proposes The evidence offered is either self-reported (BenchSci's own '22% of key projects found novel targets', Causaly's own customer counts) or methodological rather than outcome-based (Standigm's ASK paper in Briefings in Bioinformatics, 2024, with no partnered clinical asset behind it). Caveat: this radar records traction and rounds, not every validation study a vendor may have published, so the absence is of a commercial product built on audited hit rates in this snapshot — and note that outcome prediction for assets already in trials is a different job that sits in the radar's 'Sponsor-side trial design, digital twins & translational AI' segment (Intelligencia AI, Unlearn.AI), not here. + Gap No row indexes failure — discontinued programmes dead targets, negative results — as a first-class asset Every mining engine described here extracts positive causal assertions or activity data: Causaly's causal graph, Biorelate's Galactic AI over papers and patents, Excelra's GOSTAR structure-activity records, Genomenon's variant evidence sets. Caveat: the richest failure data is proprietary to pharma and unpublished, which may be the reason rather than an oversight, and ONTOFORCE's DISQOVER does fuse a customer's internal data with ~70 public sources — the row simply does not claim negative-result coverage. Absence is stated about this radar's twelve rows, not about the market at large. + Gap No vendor here couples its graph to a validation loop and prices on validated targets The two halves exist on this radar but are not joined: Relation runs the loop internally (patient tissue, perturbation, MORGAN) and monetises it as partnered discovery, not as a product, while wet-lab validation as a purchasable service is a separate segment of this radar entirely (Adaptyv Bio, LabGenius, Arctoris, Tierra Biosciences, Nuclera) with no knowledge-graph vendor attached to it. Everyone else sells seats, licences or data. Caveat: outcome-based pricing shifts scientific risk onto the vendor and may be commercially unattractive rather than unoccupied; the claim is that no row in this segment does it, not that it has never been tried. + Gap Every named buyer in this segment is a pharma R&D or research-centre account — Novo Nordisk, Novartis, Takeda, Ipsen, J&J, Gilead, Regeneron, the FDA and NIEHS (Causaly); Lilly, Sanofi, Moderna, AstraZeneca (BenchSci); AstraZeneca, Amgen, Princess Maxima Center (ONTOFORCE); Eisai, Daiichi Sankyo, UBE (FRONTEO). Nobody sells the same evidence graph to adjacent buyers who ask the same question with different money behind it: investor and BD diligence on a target's prior art, payers, or disease foundations. Caveat: those buyers may simply lack the budget line, and the radar's investor segment lists capital pools (Dimension, ARCH, Flagship, DCVC Bio) rather than tooling vendors, so nothing on this radar contradicts the gap — but the gap is asserted only for this snapshot. + Gap No row sells the provenance and verification layer for agent-generated biomedical claims — the auditable chain from a stated assertion back to the licensed source that supports it. That this is now the binding constraint is visible in Causaly's July 2026 Sage deal buying agents full-text access to 400+ journals: the answer to 'can the agent read the paper' was a publishing contract, not a product anyone here sells. Caveat: citation-grounding may be built inside these platforms without appearing in a traction row, and Genomenon's Mastermind already supplies gene/variant evidence used in regulatory dossiers — the unoccupied job is the standalone verification layer across vendors, in this radar's snapshot. S9 Omics & single-cell models open 4 5 gaps identified — open for the gaps, incumbents and entry risks. Immunai US · Generates and assembles single-cell immune profiling data at scale and trains AMICA, an immune-system foundation model, then sells the resulting platform to pharma to pick patients, targets and combinations in clinical development. Bioptimus EU · Builds multi-modal foundation models of biology — H-Optimus for pathology slides, M-Optimus spanning several biological modalities — and assembles the paired clinical/molecular corpus to train them. Synthesize Bio US · Sells GEM, a generative model of bulk RNA-seq that produces in-silico expression data from a written experimental design, aimed at pharma teams sizing experiments before running them. CytoReason Asia · Builds computational disease models on an evidence graph of molecular and clinical data, sold to pharma as a subscription for target selection and trial design. + Gap Nobody here sells a perturbation-response or virtual-cell model — predicting how a cell shifts under a genetic or drug perturbation. The four rows model what exists (Immunai profiles immune states, CytoReason assembles an evidence graph, Bioptimus trains on archival specimens) or simulate a described experiment (Synthesize Bio's GEM), none predicts an unseen intervention. Caveat: this radar's data-generation half is populated elsewhere — Tahoe Therapeutics and Somite AI in 'Cell-programming foundries & lab-generated training data', Adaptyv Bio and Arctoris in 'Closed-loop wet-lab validation' — so the missing piece is the model layer over that data within this 174-row August 2026 snapshot, not in the market at large, where academic and open efforts fall outside the radar's funded-company cut. + Gap No independent evaluation of omics foundation models exists anywhere in this radar Every traction line in this segment is a contract, a corpus size or a self-published result — CytoReason's top-20-pharma reach and Synthesize Bio's 'generated data matching wet-lab results' are both the vendor's own account — and the nearest evaluation-shaped segment, 'Sponsor-side trial design, digital twins & translational AI' (Unlearn.AI, Intelligencia AI, QuantHealth), scores trials rather than biology models. Caveat: absence is from this snapshot only; consortium and academic benchmarks would not appear as rows in a company radar. + Gap Single-cell modalities beyond transcriptome and image are unclaimed here Immunai's corpus is single-cell immune profiling, Synthesize Bio is bulk RNA-seq, Bioptimus is pathology slides plus a multi-modal line — no row builds a proteomic or metabolomic single-cell foundation model. Caveat: proteomics is present in this radar but as a data asset, not a model — Protai sits in 'Proprietary biological data assets for model training' — so the gap is the model over the modality, and only within this snapshot. + Gap There is no self-serve, metered way to buy single-cell inference below top-20-pharma scale Bioptimus is the sole row with non-bespoke distribution (open weights, then AWS Marketplace); Immunai and CytoReason reach customers only through negotiated multi-year alliances, and Synthesize Bio has disclosed no customer at all. A mid-cap biotech or an academic group appears in nobody's go-to-market. Caveat: a low-priced self-serve vendor may simply sit below the funding and press threshold at which this radar picks companies up, so read this as unclaimed by the radar's 174 named companies, not as proven market absence. + Gap No row here carries an omics model into regulated clinical deployment — a cleared diagnostic or a companion-diagnostic path. All four stop at pharma R&D decision support: patient, target and combination selection (Immunai), target selection and trial design (CytoReason), experiment design (Synthesize Bio), model supply (Bioptimus). Caveat: this is partly an artifact of how the radar is cut — the regulated-deployment names would sit in 'Computational pathology & spatial-biology models' (Proscia, Deciphex, Indica Labs), an image-based segment whose clearance status this radar does not record — so the honest claim is that no row in THIS segment shows a regulatory path, not that none exists. S10 Pathology & spatial biology contested 12 5 gaps identified — open for the gaps, incumbents and entry risks. Aignostics EU · Builds pathology foundation models (RudolfV, Atlas, Atlas 2) that read whole-slide tissue images, and sells the resulting spatial-biology readouts to pharma for biomarker discovery and translational research. Noetik US · Trains virtual-cell foundation models (OCTO-VC) on its own paired pathology-plus-spatial-transcriptomics data to simulate tumour tissue in silico. Proscia US · Concentriq, the image-management and analysis platform pharma and labs run their whole-slide pathology on, plus its own AI applications. Deciphex EU · Patholytix, AI-assisted preclinical/toxicologic pathology review sold to sponsors and CROs (plus Diagnexia, a human subspecialist reading network). Mindpeak EU · Deep-learning IHC and H&E readers (breast, prostate markers) that score tissue automatically for labs and pharma studies. Nucleai US · Spatial AI biomarkers read off multiplexed tissue images to pick which patients a sponsor's drug will work in. Modella AI US · Multimodal pathology foundation models and agents (PathChat, Judith) for oncology biomarker discovery and clinical development. HistoIndex Asia · Stain-free second-harmonic-generation imaging plus the qFibrosis algorithm to quantify liver and other fibrosis in trial biopsies. PharmaNest US · FibroNest digital-pathology biomarkers that score fibrosis and inflammation on trial biopsies for sponsors across MASH, IBD, EoE and other fibrotic indications. AIRA Matrix Asia · Deep-learning image analysis for drug discovery and preclinical/toxicologic pathology studies, sold to pharma and CRO research groups. Enable Medicine US · Generates, stores and analyses spatial-biology data on one cloud platform so biopharma can ask which cell neighbourhoods drive response or resistance. Indica Labs US · HALO / HALO AI quantitative image analysis and HALO AP, used by pharma services teams for biomarker scoring and companion-diagnostic development. + Gap Nobody here sells per-case clinical diagnostic reading against a reimbursement code Every row monetises pharma R&D budgets (Aignostics, Noetik, Nucleai), platform seats (Proscia, Indica Labs) or preclinical safety studies (Deciphex, AIRA Matrix), and the segment's only FDA clearance is Leica/Indica's Dec 2025 enterprise platform, not an algorithm; HistoIndex ships FibroSIGHT Plus as an LDT by choice. Caveat: this radar is scoped to AI for bioscience and pharma and no other of its twenty segments carries a clinical-diagnostics vendor either, so this is an absence from this snapshot's frame, not evidence the job is unserved in the market. + Gap No independent evaluator of pathology models appears anywhere in this radar The need is visible in the rows' own hedges — Aignostics' 80-benchmark result is company-reported, Indica's 200% growth is vendor-stated, Mindpeak's 30,000+ diagnoses is a vendor figure — and a check of the corpus index finds no benchmarking or audit house in any of the twenty segments, including Scientific data infrastructure and Lab systems of record. Caveat: the index gives me company names, not business descriptions, so a firm listed in another segment could run third-party evaluation as a side line I cannot see from here. + Gap Nothing in this segment spans tissue readout through to a designed intervention Noetik simulates tumour tissue in silico and licenses the model to GSK but has no drug in the clinic, and the molecule-design work sits in other segments of this same radar (AI-native therapeutics preclinical, Computational chemistry) — so this is a seam nobody straddles, not a job the radar is missing. Caveat: a pharma partner may be closing that loop internally where the radar cannot observe it, and the GSK anchor and the AstraZeneca acquisition of Modella AI are exactly the shape that would. + Gap Tissue outside oncology and liver/GI fibrosis is unrepresented in this snapshot The named indications run oncology (Aignostics, Noetik's NSCLC and CRC models, Nucleai, Mindpeak's breast and prostate markers, Modella AI) and fibrosis (HistoIndex on MASH, PharmaNest across MASH, IBD and EoE); no row names neurology, nephrology, cardiac or transplant-rejection pathology. Caveat: Proscia, Indica Labs and AIRA Matrix describe general-purpose image-analysis platforms that may already run those indications without the row saying so. + Gap No row sells its pathology model as infrastructure other builders rent Every foundation-model row here — Aignostics, Noetik, Modella AI before the acquisition — monetises through a bilateral pharma agreement, and the two platform rows (Proscia, Indica Labs) ship their own AI rather than hosting anyone else's; there is no API, per-inference tier or self-serve path described in twelve rows. Caveat: this is an absence in an August 2026 snapshot assembled from funding and partnership announcements, and a developer tier is precisely the kind of launch that would not generate the press release these rows are sourced from. S11 Autonomous labs open 4 5 gaps identified — open for the gaps, incumbents and entry risks. Lila Sciences US · Runs 'AI Science Factories' - robotic labs where models propose hypotheses, execute experiments and choose the next round - across life, chemical and materials sciences. Automata UK · Sells LINQ, a modular robotic bench plus cloud software that chains third-party instruments into end-to-end automated lab workflows. Atinary Technologies US · No-code Bayesian-optimisation software (SDLabs) that picks the next experiment and drives robotic hardware for chemistry and materials campaigns. Tetsuwan Scientific US · Builds an LLM-driven robotic scientist that translates a research goal into executed bench protocols on off-the-shelf liquid handlers. + Gap Nobody in this segment publishes an audited measure of whether autonomy actually beats a well-run human lab — experiments per week, hit rate, cost per validated result. Evidence: Lila's row says outright that capital raised is the datapoint and no discovery output has been disclosed; Automata's customer count is vendor-stated and Atinary's flagship deployments reach us through an AWS case study. Caveat: this is an absence in this radar's four-row snapshot, not proof no such benchmark exists — a pharma or academic consortium could be measuring this privately, and the nearest radar neighbours in 'Closed-loop wet-lab validation as a service' (Adaptyv Bio, LabGenius, Arctoris) sell validation runs rather than third-party audits of anyone's autonomy claims. + Gap No row carries an autonomous loop into a regulated GxP-validated setting — every one is discovery-side (Lila's factories, Atinary's chemistry and materials campaigns, Tetsuwan's research assays). Check the corpus first, though: regulated work is not missing from this radar, it sits in other segments — 'AI for GxP biomanufacturing operations' (Aizon, Apprentice.io, Basetwo) and 'Scientific data infrastructure (instrument-to-cloud, GxP)' (TetraScience, Ketryx, Artificial). So the unclaimed job is specifically the bridge: letting a model choose the next experiment inside a validated environment. Caveat: that bridge may be a deliberate avoidance rather than an opening, since qualifying a self-selecting system under GxP is a genuinely unsolved regulatory question. + Gap A self-driving loop aimed squarely at cell and tissue biology has no funded owner here Atinary is explicitly chemistry and materials; Lila spans life, chemical and materials sciences but discloses nothing about which; Tetsuwan is the only life-science-specific loop and it is a $5.7M pre-seed company with one deployment. Caveat and cross-reference: wet biology at scale is well covered elsewhere in this radar — 'Cell-programming foundries & lab-generated training data' (Ginkgo Bioworks, Tahoe Therapeutics, Somite AI, Asimov) and the wet-lab-validation segment run enormous biological throughput. What none of them is, on this snapshot, is closed-loop and model-directed; they run campaigns a human designed. The gap is partly a segment-boundary artifact, so treat it as a hypothesis about this radar's coverage rather than a market fact. + Gap Nobody rents the loop by the experiment All four business shapes require the buyer to commit: build and own the facility (Lila), buy the bench (Automata), license the software and supply your own robots (Atinary), or have a system installed in your lab (Tetsuwan, one install so far). Caveat and cross-reference: remote-access lab work does exist in this radar — Arctoris and Adaptyv Bio in 'Closed-loop wet-lab validation as a service' run experiments for you — so the absent thing is not remote execution but remote access to the decision layer, paying per campaign for a model that picks what runs next. Honest limit: with four rows this may simply be a shape the radar did not capture rather than one nobody has tried. + Gap The exhaust of autonomous runs — including the failures — is not being sold by anyone in this segment Lila keeps its factories in-house and discloses no output; Atinary and Tetsuwan run inside customer facilities (Takeda, MIT, La Jolla Labs), so the traces stay with the customer; Automata's row describes workflow chaining, not data products. Caveat and cross-reference: lab-generated training data is a real business elsewhere in this radar — 'Proprietary biological data assets for model training' (A-Alpha Bio, Variant Bio, deCODE genetics) and the cell-programming foundries — so the specific unowned job is the negative-result and dead-end trace that only a self-driving lab generates at volume. Whether that data has buyers is untested by anything in these rows. S12 Wet-lab validation contested 5 5 gaps identified — open for the gaps, incumbents and entry risks. Adaptyv Bio US · Cloud protein foundry: designers submit sequences through an API, Adaptyv expresses and assays them, and returns binding/activity data to retrain the model. LabGenius UK · Runs EVA, a robotic design-build-test loop that proposes antibody variants, makes and screens them, and retrains on its own assay data. Arctoris UK · Operates Ulysses, a fully robotic wet-lab that runs biology assays as a service for AI drug-discovery partners and generates structured DMTA data. Tierra Biosciences US · Makes and tests customer-designed proteins on demand using cell-free synthesis, returning both the protein and the expression data. Nuclera UK · Sells eProtein Discovery, a benchtop digital-microfluidics system that expresses and purifies protein constructs in under 48 hours. + Gap Nobody here sells cell-based or in-vivo functional validation as a service All five rows stop at protein expression, purification or binding (Adaptyv returns binding/activity, Tierra returns protein plus expression data, Nuclera expresses and purifies constructs in under 48 hours), and Arctoris is the only general-biology assay operator but its named closed loops are with IBM Research and SpiroChem on design-make-test chemistry, with no cell-phenotype offering described. Caveat: this is an absence in THIS radar's wet-lab-validation segment only, and the adjacent job of generating cell-perturbation data does have owners elsewhere on the radar in Cell-programming foundries and lab-generated training data (Ginkgo Bioworks, Tahoe Therapeutics, Somite AI). + Gap No independent, audited benchmark of design-model hit rates Every traction figure in the segment is self-reported: Adaptyv's protein count and 30+ organisations are marked vendor-stated, Tierra's AI capability is company-described and not benchmarked publicly, LabGenius names no candidate from Sanofi. Adaptyv's public design competitions are the only independently visible artefact and they are its own marketing surface, not a neutral referee. Caveat: no segment in this radar's 20 occupies a referee role, so this is an absence across the snapshot as a whole, not evidence that no academic or nonprofit bench exists outside it. + Gap No capacity broker or multi-foundry routing layer Each row is one vertically integrated site selling only its own throughput: Arctoris one Oxford lab (capacity roughly doubled by absorbing Eli Lilly's San Diego automation lab), Tierra one San Leandro cell-free operation, Adaptyv one API-fronted foundry. Nothing in the rows routes a design job to whichever foundry is cheapest or fastest, and Nuclera, the best-capitalised distributor here, ships hardware rather than brokered capacity. Caveat: absence in this radar's snapshot of five rows, which is small enough that one unlisted broker would overturn it. + Gap Nobody sells the negative data Every row's pitch is returning data to the customer who commissioned it (Adaptyv returns results to retrain the customer's model, Tierra returns expression data, Arctoris generates structured DMTA data for partners); no row describes pooling failed designs across customers into a shared training corpus, which is the asset a validation service is uniquely positioned to accumulate. Caveat: the proprietary-data-asset business model is on this radar, just not attached to a validation service here, in Proprietary biological data assets for model training (A-Alpha Bio, Protai, Variant Bio), so this is a gap in this segment's snapshot rather than an unclaimed idea. + Gap No regulated-grade validation output Everything described in these rows is research-grade discovery data, and nothing in the segment claims a GxP-compliant, audit-trailed or regulatory-submittable package from the loop, even though Arctoris and Nuclera both sit close to pharma buyers who eventually need one. Caveat: GxP is covered twice elsewhere on this radar, in Scientific data infrastructure (instrument-to-cloud, GxP) (TetraScience, Ketryx) and AI for GxP biomanufacturing operations (Aizon, Apprentice.io), so the missing piece here is the join between a closed-loop assay service and those compliance layers, not GxP capability in the market. S13 Foundries & training data open 5 5 gaps identified — open for the gaps, incumbents and entry risks. Ginkgo Bioworks US · Runs an automated cell-engineering foundry and sells lab-generated biological training datasets and assay panels (Datapoints), cloud-lab access and bioprocess services to other people's AI model builders — picks-and-shovels rather than a model builder. Tahoe Therapeutics US · Industrialises single-cell perturbation experiments and sells or open-sources the resulting datasets as training data for virtual-cell models. Somite AI US · Builds DeltaStem, a foundation model for human stem-cell differentiation, trained on signalling data its own capsule platform generates. Asimov US · Designs mammalian cell lines and genetic circuits for biologics manufacturing, packaging them as licensable engineered genetic systems. Bota Biosciences Other · Industrial-biotech foundry combining lab automation and computational design to develop enzymes and fermentation routes for chemical manufacturers. + Gap Nobody here certifies the data they sell Ginkgo ships ADME-One panels and Tahoe ships Tahoe-100M, but no row offers batch-effect, reproducibility or provenance verification of a dataset sold to a third party — the buyer takes the vendor's word. Check the index before calling this missing outright: the adjacent capability sits in another segment of this radar, Scientific data infrastructure (TetraScience, Elucidata, Ganymede Bio, Ketryx), which handles a customer's own instrument-to-cloud and GxP pipeline rather than auditing a purchased training set. Caveat: this is an absence in the radar's August 2026 snapshot of these five rows, not evidence that no such service exists in the market. + Gap No pricing or licensing layer between data producer and model builder The two revenue shapes visible are direct catalogue sale (Ginkgo at $199 per panel) and free release into a nonprofit atlas (Tahoe into Arc's Virtual Cell Atlas) — no broker, no royalty or downstream value-share where the data producer participates in the model it trained. Caveat: no row in any of this radar's 20 segments is described as a data marketplace or licensing exchange, but the radar records only 174 companies and this claim is about that snapshot, not about the market. + Gap No foundry here runs perturbations in primary human tissue or in vivo Tahoe's 100M+ profiles come from 50 immortalised cancer lines, Somite's from its own stem-cell capsule platform, Ginkgo's from assay panels, Asimov's from engineered mammalian production lines — all cultured systems. The radar's human-derived data supply sits in a different segment, Proprietary biological data assets for model training (deCODE genetics, MedGenome, Nightingale Health, Variant Bio, Protai, A-Alpha Bio), so the gap is specifically a foundry that *perturbs* primary human material at industrial scale. Caveat: snapshot-level only, and rows in other segments were not read in detail here. + Gap No disclosed commercial supply contract with a named AI-model builder The picks-and-shovels thesis needs a visible buyer, and the only named counterparties in these rows are research institutions — PNNL, MIT, Caltech, University of Maryland, Northwestern for Ginkgo; Arc Institute and CZ Biohub for Tahoe. None of the radar's seven Biomolecular foundation model rows (Chai Discovery, EvolutionaryScale, Profluent Bio, Basecamp Research, BioMap, Helixon, Deep Genomics) appears as a customer in any row here. Caveat: absence of a *disclosed* deal is not absence of a deal — supply contracts are routinely unannounced — so this is a gap in what the snapshot can see, not a proven commercial vacuum. + Gap No EU-domiciled foundry in this segment The five rows are Boston (Ginkgo), San Francisco (Tahoe), Boston (Somite), Cambridge MA (Asimov) and Hangzhou with a Lafayette, California second site (Bota) — so a European pharma customer with data-residency or supply-sovereignty constraints has no in-region cell-programming foundry in this snapshot. Caveat: this is a claim about these five rows only; other segments of the radar were not inspected for HQ, and the radar is not a census of the market. S14 Scientific data infrastructure contested 12 5 gaps identified — open for the gaps, incumbents and entry risks. TetraScience US · Runs a scientific data pipeline — pulls raw output off lab instruments, parses and harmonizes it into a common model, and lands it in cloud warehouses so pharma can run analytics and train models on it; now marketed as 'Tetra OS, the operating system for scientific intelligence'. Elucidata UK · Harmonizes messy public and internal multi-omics datasets into a single machine-readable model on its Polly platform, so pharma teams can train models on biological data without rebuilding curation pipelines each time. Synthace UK · Digital-experiment platform — biologists design multi-factor experiments in a GUI, simulate them, execute them on their own liquid handlers, and get structured data and metadata back automatically. Ganymede Bio US · Cloud IDE that lets scientists write code-level integrations from any lab instrument or scientific app into a governed data warehouse, with a GxP-validated software lifecycle on top. Scitara US · Middleware (Digital Lab Exchange) that connects instruments, ELN/LIMS and cloud apps into one regulated-market-compliant integration fabric, sold as connectivity rather than a system of record. Seqera EU · Commercial layer on Nextflow — orchestrates and reproducibly runs large bioinformatics pipelines across clouds and HPC for genomics and multi-omics teams. Lifebit UK · Federated trusted research environment — lets pharma and researchers run analyses against national biobank and health datasets without the data ever leaving the custodian's jurisdiction. DNAnexus US · Regulated cloud for genomic and multi-omic data — storage, pipeline execution and collaboration for biobank-scale cohorts. Velsera US · Bioinformatics and clinicogenomic data platform (Seven Bridges) plus variant-interpretation software, sold to pharma, diagnostics labs and cancer research consortia. Zifo RnD Solutions Asia · Scientific-informatics services firm that implements and integrates R&D, QC and manufacturing data systems for pharma, biotech and device companies — the delivery arm behind many GxP lab-data programmes. Artificial US · aLab Suite — orchestration software that schedules and drives robotic lab equipment and guides human bench steps, sitting between the scheduler and the instruments. Ketryx US · Compliance automation for regulated life-sciences software — keeps requirements, code, tests and evidence continuously audit-ready so AI and software components can ship under FDA/GxP rules. + Gap Nobody here sells the capture layer that sits on the instrument itself, on the OEM's side of the wire Evidence: Thermo Fisher appears in this radar only as a TetraScience co-marketing partner (12 Jan 2026), and no instrument maker is a row in ANY of the twenty segments in the corpus index — every row here starts downstream, pulling data off equipment somebody else built. Caveat: this is an absence in THIS RADAR's August 2026 snapshot, which tracks venture-scale AI-native names; large diversified instrument vendors are plausibly doing exactly this work and are simply out of the radar's scope, so treat it as unobserved, not unoccupied. + Gap Nobody owns the seam between an autonomous lab and the GxP warehouse Evidence: Artificial drives robotic equipment and Synthace designs and executes experiments, but both sell into the lab, not into the regulated data estate; meanwhile the corpus index carries whole separate segments for 'Autonomous labs & self-driving experimentation' (Lila Sciences, Automata, Atinary, Tetsuwan Scientific) and 'Closed-loop wet-lab validation as a service' (Adaptyv Bio, LabGenius, Arctoris, Tierra, Nuclera) — so the labs exist in this radar, they just have no counterpart here that lands their exhaust into a validated, audit-ready store. Caveat: the seam may be covered inside those companies' own stacks in ways the traction summaries do not surface. + Gap Nobody sells proof that harmonized data actually improves a model Evidence: every row's traction is measured in inputs — datasets (Elucidata's ~1.5M), petabytes (DNAnexus's 65+), records (Velsera's 175M+), deployments (TetraScience's Bayer, Organon, Syngenta, Takeda) — and not one names a downstream model-performance result; nor is any segment in the corpus index framed as evaluation or benchmarking of scientific data quality. Caveat: absence of the claim in a funding-press-derived snapshot is weak evidence of absence of the capability, and buyers may simply be measuring this privately. + Gap Nobody sells the exit getting a pharma's harmonized data OUT of one platform and into another. Evidence: all twelve rows describe ingestion into their own model or environment, and none mentions export, portability or a neutral interchange format — with TetraScience's earliest named deployments now multi-year (the Bayer relationship was an expansion as of 20 Nov 2025) there is a re-platforming job forming that has no vendor. Caveat: Scitara sells an integration fabric that moves data between live systems, which is adjacent; whether that constitutes an exit path cannot be judged from these rows, and 'lock-in' is an inference I cannot verify here. + Gap Nobody in this segment's snapshot targets a buyer below large pharma and national genome programmes Evidence: the named customers are Bayer, Organon, Takeda, Syngenta, Pfizer, Genentech, Janssen, Genomics England, the NIH, Singapore's Ministry of Health and the Danish National Genome Center — enterprise and sovereign accounts throughout; Ganymede Bio's early-stage-biotech commitments are the only downmarket signal and are a vendor claim of ~$1M. Caveat: Seqera's Nextflow base is a de facto academic standard, so cheap academic adoption almost certainly exists but is invisible in press-release traction — this is an absence in what THIS RADAR records, not a claim that small labs are unserved. S15 Lab systems of record saturated 7 4 gaps identified — open for the gaps, incumbents and entry risks. Benchling US · Sells the cloud system of record for life-science R&D — electronic lab notebook, sample/inventory registry and results capture — and layers AI agents (Ask, Deep Research) and MCP-based connectors on top of that data. Dotmatics US · Sells the incumbent stack of scientific R&D software — ELN, compound/biologics registration, assay data management and GraphPad Prism — to pharma and biotech discovery teams. Sapio Sciences US · Single-platform LIMS + ELN + scientific data management with a built-in knowledge graph and 200+ instrument integrations, aimed at biopharma discovery and QC labs. Scispot Other · No-code LIMS/ELN/SDMS for biotech and diagnostics labs, positioned as the context and execution layer that AI agents call into to read lab state and trigger lab actions. SciSure US · Combined ELN, inventory, LIMS and lab safety/compliance platform for academic and biotech research organisations — a system of record that spans science and EHS. Cenevo UK · Parent of Labguru (cloud ELN/LIMS) and Mosaic (sample and compound management) — the record of what was done and where every sample sits, now marketed with agentic AI for protocol conversion and inventory search. L7 Informatics US · Enterprise Science Platform — unifies LIMS, ELN and workflow orchestration across therapeutics, diagnostics and cell/gene manufacturing, so a sample's process and its data live in one system. + Gap Nobody sells escape from the record -- extraction schema mapping and migration off an incumbent ELN/LIMS -- even though this segment just manufactured the demand: Cenevo merged Labguru with Titian's Mosaic, and SciSure absorbed eLabNext, then Labfolder and Labregister, leaving customers on assembled stacks nobody chose whole. Caveat: absent from these 7 rows only, and the radar's 'Scientific data infrastructure (instrument-to-cloud, GxP)' segment holds the nearest neighbours (TetraScience, Scitara, Ganymede Bio) -- they move instrument data to the cloud rather than lift a lab off its system of record, so one of them could extend into this before a new entrant does. + Gap No row publishes an accuracy, eval or audit measure for the AI that answers over the lab record Benchling ships Ask and Deep Research plus MCP connectors, Cenevo markets agentic protocol conversion, Sapio announced an NVIDIA BioNeMo integration -- and not one reports a correctness number or AI-attributable revenue; the sharpest quantitative AI claim in the whole segment is a Benchling exec forecasting that 75% of biology data-analysis tasks will be handled by agents within a year, which is a prediction, not a result. Caveat: this is the absence in this radar's 7-row snapshot -- Ketryx in 'Scientific data infrastructure' is the closest thing the radar has to a validation/compliance vendor, and the eval work may be unpublished rather than undone. + Gap Nobody here is a bottom-up, self-serve, per-seat record a two-person lab can buy without a sales cycle The rows are enterprise or institutional by construction: Sapio names Bristol Myers Squibb, GSK, J&J, Charles River, LabCorp and the Wellcome Sanger Institute; Cenevo claims 8 of the top 10 biopharma; Benchling's 7,500+ academic institutions are reached institutionally; SciSure's academic base is tied to EHS/safety compliance. Scispot is closest and still sells to '100+ enterprise labs'. Caveat: this describes the 7 funded companies in this radar, which is a capital-weighted snapshot -- a self-serve or open-source ELN could sit below the funding threshold that got a company into the corpus at all. + Gap Nobody owns the record as a transactional surface an agent can write to with guarantees -- permissions, rollback, and provenance on machine-initiated changes to samples and results. Scispot is alone in even positioning that way ('the context and execution layer that AI agents call into to read lab state and trigger lab actions'), and Benchling's Apr-2026 AI Connectors expose data to Claude/ChatGPT via MCP, which is the read side. Caveat: the write side of autonomous execution sits in this radar's 'Autonomous labs & self-driving experimentation' segment (Lila Sciences, Automata, Atinary Technologies, Tetsuwan Scientific) -- they own the robot, not the system of record, so the gap is the seam between the two rather than a category missing from the market. S16 Biomanufacturing operations contested 5 5 gaps identified — open for the gaps, incumbents and entry risks. Culture Biosciences US · Sells cloud-connected bioreactors (Stratyx 250) and the Culture Console software that runs process-development experiments remotely, for both mammalian and microbial scale-up. Aizon US · GxP-compliant manufacturing software for pharma and biopharma plants - data historian, electronic batch records, predictive process analytics, and a newer agentic layer. Apprentice.io US · GxP manufacturing cloud (MES, LES and eLogs) for batch pharma, now layering predictive and agentic AI on top of the execution record. Basetwo Other · Physics-informed AI models of pharma and chemical unit operations that let process engineers optimise and eventually auto-control production. Differential Bio EU · Virtual scale-up platform pairing miniaturised parallel microbiology with ML to predict how a fermentation process behaves at production scale. + Gap Autonomous closed-loop control of a regulated production line that has actually shipped Basetwo's AutoPilot is 'announced, not shipped'; Aizon's agentic upgrade was pre-announced in Oct 2025 for early-Q1-2026 with shipped status unverified in the row; Culture's Gemini work is a three-phase plan. Three of five rows point at the same destination and no row in this radar's snapshot evidences arriving. Caveat: this is an absence in the snapshot's sources as of the radar vintage, not proof nobody has shipped since - and the likeliest cause is the validation path, so the gap may be regulatory rather than technical. + Gap A cross-plant or federated process model that improves with every customer Every deployment described here is single-site: Aizon's proof is one Recordati line, Basetwo names no customer at all, and Apprentice's headline metric (>30M requests/month) counts traffic rather than model performance - no row in this segment claims a data network effect. Caveat: pooling manufacturing data across competing sponsors is commercially and legally hard, so the absence may be a market fact rather than an unclaimed opportunity, and the rows describe shipped products, not roadmaps. + Gap Independent, audited verification of manufacturing-AI performance claims Every efficiency figure in this segment is vendor-supplied and the rows say so - Culture's 16% / 25% / 30%, Basetwo's 40% and 25%, Aizon's 1.5% at Recordati - and no row sells third-party measurement of anyone else's numbers. Caveat: I checked the corpus index and no other segment of this radar covers benchmarking or verification either, but a radar of funded startups would not capture consultancies, notified bodies or CROs doing this work off-radar. + Gap Turning the execution record into regulatory output - CMC submission sections, deviation investigations, CAPA narratives - rather than into process optimisation. Apprentice sells MES, LES and eLogs; Aizon sells electronic batch records plus predictive analytics; both hold the compliance data and neither row claims it generates the filing or the investigation. Caveat: this radar routes GxP data plumbing to a separate segment (Scientific data infrastructure - TetraScience, Ganymede Bio, Ketryx and others), so check those rows before treating the job as empty; and the absence is read from five 'what' descriptions, which may omit a shipped feature. + Gap A low-touch product for an operator that cannot buy an enterprise MES All five rows require scale: Apprentice and Aizon land large named manufacturers, Culture requires physical bioreactor capacity (Stratyx 250), and the two model-first rows are pre-commercial or customer-less. Caveat: pricing and go-to-market appear nowhere in the rows, so this is inferred from customer shape; and the small-batch, per-patient variant of the job sits in this radar's Cell-therapy manufacturing automation segment (Cellares, Ori Biotech, Multiply Labs and others), so that case is covered elsewhere here rather than missing. S17 Cell-therapy manufacturing contested 6 5 gaps identified — open for the gaps, incumbents and entry risks. Cellares US · Operates as an 'IDMO' - builds the Cell Shuttle, an end-to-end automated cell-therapy manufacturing box, and runs it as a contract manufacturing service in its own facilities. Cellino US · Automates patient-specific (autologous) stem-cell manufacturing using label-free imaging plus machine-learning cell classification that drives a high-speed laser to remove unwanted cells in a closed cassette. Ori Biotech UK · Sells IRO, a closed automated platform that runs the full CAR-T and advanced-therapy manufacturing process on one box. Cellular Origins UK · Constellation, a mobile-robot plus sterile-welding system that links existing bioprocess instruments so a cell therapy scales without redevelopment. Multiply Labs US · Robotic clusters of collaborative arms that operate standard cell-therapy instruments inside a GMP suite, replacing manual manipulation. CellFE US · Infinity MTx, a microfluidic non-viral gene-delivery system that automates the engineering step of cell-therapy manufacture. + Gap Nobody in this segment sells the process-control intelligence as a product detached from their own metal Evidence: five of six rows are peripheral-or-marketing on aiRole, and the one core-AI row (Cellino) uses its ML only to drive its own laser inside its own cassette — no row offers a model that improves a run on someone else's instrument. Caveat: this is absence from THIS radar's snapshot of this segment only; the adjacent 'AI for GxP biomanufacturing operations' segment (Aizon, Apprentice.io, Basetwo, Culture Biosciences) and 'Scientific data infrastructure' (TetraScience, Ganymede Bio) hold the general bioprocess-AI and instrument-to-cloud players, so the gap is specifically cell-therapy process intelligence sold standalone, not process AI as a category. + Gap No row automates the decision that a batch passes — potency, identity, sterility, predictive release Evidence: every 'what' field describes making the product (Cellares' end-to-end box, Ori's full-process platform, Cellular Origins' instrument linking, Multiply Labs' robot arms, CellFE's transfection step); none mentions QC, assay, or release testing. Caveat: absence from this radar's six rows and their short descriptions — a row could run release assays without disclosing it in the cited source, and analytics-heavy players sit in other segments of this radar (Computational pathology & spatial-biology; Scientific data infrastructure). + Gap Nobody here is building for decentralised point-of-care manufacturing at the hospital Evidence: all six rows are centralised-facility bets — Cellares manufactures in its own facilities, Multiply Labs deploys into Stanford's GMP suite and works with UCSF, Cellular Origins links instruments inside an existing plant. Caveat: this is a claim about the six rows in this radar, not the market; Cellares' IDMO model is an explicit bet the other way (buy capacity centrally), so the absence may reflect a deliberate industry position rather than an unclaimed job. + Gap No row sells the vein-to-vein orchestration layer — chain of identity, scheduling, and logistics around an autologous batch. Evidence: all six rows sell hardware or a service running hardware; none of the descriptions covers patient-to-patient tracking or slot scheduling, even though Cellares' $380M agreement is structured as reserved capacity, which is a scheduling problem. Caveat: absence from this segment in this radar only; the 'Lab systems of record (ELN / LIMS)' segment (Benchling, L7 Informatics, Dotmatics) and 'Scientific data infrastructure' hold adjacent software, though no row in either is described as cell-therapy vein-to-vein orchestration. + Gap There is no independent, audited benchmark of what any of these platforms actually costs or yields Evidence: the segment's single quantified economic claim is Multiply Labs' 74% cost reduction, and the row itself says it is a published UCSF result for one process, not a commercial cost at scale; every other row offers designations, partnerships or placements instead of numbers. Caveat: a third-party benchmarking body could exist outside this radar's 174 rows, and the Multiply Labs figure is peer-reviewed, so it is stronger evidence than a pure vendor claim — the gap is comparative benchmarking across platforms, not measurement per se. S18 Industrial biology contested 10 4 gaps identified — open for the gaps, incumbents and entry risks. Arzeda US · Computationally designs enzymes and proteins for industrial production, and runs cell-free enzymatic manufacturing of specialty ingredients. Antheia US · Manufactures pharmaceutical key starting materials and APIs by fermentation in engineered yeast, replacing plant-extracted supply chains. Epoch Biodesign UK · Designs enzymes with ML that depolymerise nylon, PET and polyester at low temperature, and is building the plant to run them. Samsara Eco AU · Uses AI-designed enzymes to depolymerise mixed plastics, and is extending the same protein-design stack to recovering rare-earth minerals. ZYMVOL Biomodeling EU · Computational enzyme discovery and design house — simulates millions of in-silico variants and sends only the shortlist to the bench. bitBiome Asia · Mines a 3-billion-sequence environmental microbial genome database with ML to discover and engineer enzymes, pathways and production strains. Bluepha Other · Synthetic-biology manufacturer of PHA bioplastics and other molecules, running precision-fermentation platforms with automation and data tooling. StrainX Bioworks Asia · Full-stack precision-fermentation producer — engineers the strain, then runs downstream processing and manufacture of food proteins in-house. Pow.Bio US · Contract fermentation on a continuous, AI-controlled process that decouples growth from production, sold as capacity plus process IP. Cauldron AU · Hyper-fermentation contract manufacturer running synthetic-biology strains continuously at industrial scale. + Gap Nobody in this segment sells independent verification of fermentation or enzyme performance Four of ten rows carry traction the corpus itself flags as unaudited — Arzeda's 500 t/yr company statement, Pow.Bio's own-run 5-10x productivity figure, Cauldron's unaudited 10,000 L demonstration claim, Bluepha's disputed funding total — and no row offers third-party titer, yield or productivity attestation. Caveat: the adjacent capability sits one segment over on this radar, in AI for GxP biomanufacturing operations (Culture Biosciences runs cloud bioreactors, Aizon does GxP process analytics), so the primitive exists in the corpus even though nobody points it at industrial-biology performance claims; and this is an absence in a 174-company snapshot, not proof no such auditor exists in the market. + Gap Nobody sells the bench-to-tank bridge — a way to predict whether a designed enzyme or strain survives at production scale before the plant is built. The segment splits cleanly at that seam: the design rows (Arzeda, Epoch Biodesign, Samsara Eco, ZYMVOL, bitBiome) stop at the molecule, the operator rows (Pow.Bio, Cauldron, StrainX, Bluepha) start at the tank, and Epoch's 2028 first-production target is exactly the years that gap costs. Caveat: process modelling for regulated manufacturing lives elsewhere on this radar (Basetwo and Aizon under AI for GxP biomanufacturing operations, Atinary under Autonomous labs), so what is missing from THIS segment is anyone selling scale-up risk as a product to industrial-biology producers, not the modelling technique itself; snapshot-only claim. + Gap Nobody applies AI to downstream separation and purification, which is where much of precision-fermentation cost sits Every row is described by its strain, enzyme or reactor; StrainX Bioworks is the only row that even names downstream processing, and it does it in-house as a full-stack producer with aiRole peripheral-or-marketing. Caveat: no segment in this radar's twenty covers separations at all, so this is an absence from the whole snapshot rather than from one segment — which also means it may simply be an in-house discipline nobody has externalised into a company an AI-for-bioscience radar would catch. + Gap Nobody sells the regulatory path for novel fermented ingredients as a capability StrainX's US GRAS determination is the single food-regulatory milestone in the segment and it is a milestone StrainX won for itself; Antheia's route runs through the pharma KSM/API regime instead, which is a different file entirely. Caveat: the drug-side regulatory and trial work is represented on this radar (the therapeutics segments, and Sponsor-side trial design, digital twins & translational AI), so the gap is specifically food and industrial-ingredient dossiers; and regulatory consultancies plainly exist outside a radar that only indexes AI-native companies, so the honest claim is that no AI-native player in these 174 rows has taken the job. S19 Trial design & digital twins contested 11 5 gaps identified — open for the gaps, incumbents and entry risks. Turbine Other · Simulated Cell models the protein signalling that decides cell fate, so sponsors can run in-silico experiments on cancer and immune cell populations. Simulations Plus US · Biosimulation software for sponsors — PBPK (GastroPlus), QSP/QST and ML-based ADMET property prediction — plus modelling services. Pathos AI US · Multimodal clinical-imaging-molecular foundation model used to pick which oncology assets to run, design the trials and find the biomarkers — and to develop its own drugs. Unlearn.AI US · Builds a per-patient digital twin of the control outcome and uses it (TwinRCT / PROCOVA) to cut the number of patients a Phase 2/3 trial needs. QuantHealth Asia · Simulates a planned trial patient-by-patient before enrolment to choose dose, population and endpoints. Aitia US · Causal-AI Gemini Digital Twins of disease built from multi-omic patient data, used to nominate targets and run in-silico experiments with pharma partners and in-house. VeriSIM Life US · BIOiSIM, a hybrid mechanistic-plus-ML translational model that predicts human PK, efficacy and toxicity from preclinical data. Medidata AI US · Builds Synthetic Control Arms and simulant patients out of its historical trial repository so sponsors can shrink or replace a control group. Intelligencia AI US · Predicts probability of technical and regulatory success for an asset or indication, sold to pharma R&D portfolio teams and to investors. Median Technologies EU · Central imaging and AI imaging-biomarker services (iCRO, iBiopsy) for sponsor-run oncology, CNS and MSK trials. Genialis US · RNA foundation model (the Genialis Supermodel, trained on ~1M RNA-seq samples) that produces predictive biomarker algorithms to de-risk a sponsor's oncology trial. + Gap Nobody here sells independent, prospective validation of a simulation's accuracy Every accuracy number in this segment is self-issued — QuantHealth's 'up to 90%' and Intelligencia AI's 80% prospective / 91% retrospective, both flagged in their own rows as unpublished — and the only external check anyone holds is Unlearn.AI's EMA qualification, which certifies a method rather than auditing a track record. Caveat: this is about the radar's snapshot only. The corpus index has no benchmarking or evaluation segment anywhere across its 20 segments, so a third-party validator could exist entirely off-radar; and sponsors may prefer to validate internally rather than pay for an audit. + Gap No row runs the trial it designed Site selection, enrolment, retention and data capture are absent from all eleven rows, and Medidata AI appears here only for the Synthetic Control Arm built on its historical repository, not for trial operations. Caveat: checked against the corpus index, no other segment of this radar covers trial execution either, and Median Technologies' central-imaging iCRO work is the closest thing to it while being endpoint reading only — so this most likely reflects the radar's discovery-side scope rather than an unoccupied market. + Gap In-silico safety and toxicology has one occupant and no funded owner VeriSIM Life is alone in predicting human toxicity from preclinical data, and its row shows funding stopping at a $15M Series A in January 2022 plus an FDA-NCTR Material Transfer Agreement that the row itself calls a research collaboration, not an endorsement. Nobody in these rows packages an animal-study replacement a sponsor could actually file. Caveat: the corpus index carries no toxicology or safety-assessment segment at all, so alternatives-to-animal-testing players are outside this snapshot rather than shown to be missing from the market. + Gap Nobody carries simulated evidence into the regulatory submission itself The rows stop one step short of it — Unlearn.AI's row states qualification is not approval of any specific trial, and Medidata AI's precedent is one FDA-permitted control arm in one Medicenna Phase 3 — so dossier assembly, the statistical review package and the per-trial negotiation stay with the sponsor. Caveat: the radar puts GxP-oriented scientific data infrastructure in a separate segment (TetraScience, Ketryx and ten others), and from segment names alone I cannot rule out that one of those rows covers regulatory submissions. + Gap No disease-area franchise outside oncology Every named counterparty in these rows is a cancer programme — Turbine with MSD and AstraZeneca, Aitia with Orion and with Servier in pancreatic cancer, Genialis with Cleveland Clinic (pancreatic) and Debiopharm (WEE1), Pathos AI in mCRPC, Medidata's precedent in glioblastoma — and the two non-oncology signals are thin: Turbine's immunology deal is with an unnamed partner, and Median Technologies' CNS and MSK work is imaging services. Nothing here is built around psychiatry, rare disease or cardiometabolic trials. Caveat: within this snapshot only — Unlearn.AI and QuantHealth sell indication-agnostic methods that may already serve those areas without the radar recording it. S20 Investors & capital saturated 14 5 gaps identified — open for the gaps, incumbents and entry risks. Dimension Capital US · Techbio venture firm investing at the seam of computation and biology — the purest expression of the 'digital biology' fund thesis. ARCH Venture Partners US · Company-creating early-stage biotech firm; writes the first and often largest cheques into science-founded therapeutics platforms. Flagship Pioneering US · Venture creation firm that founds its own companies in-house rather than backing outside founders; increasingly building AI-for-science ventures directly. Lux Capital US · Frontier-science generalist fund with a substantial life-sciences book; backs computational and physical science companies well before conventional biotech investors. Sofinnova Partners EU · Europe's largest dedicated life-sciences VC, running separate strategies for therapeutics, industrial biotech and digital medicine. Novo Holdings EU · Evergreen investment arm of the Novo Nordisk Foundation — seed, venture, growth and buyout capital across life sciences, funded from a permanent balance sheet rather than fund vintages. a16z Bio + Health US · Andreessen Horowitz's dedicated bio and health practice, investing where computation, AI and biology meet. DCVC Bio US · Deep-tech life-sciences arm of DCVC; backs computational platforms that can generate many therapies or products, rather than single-asset biotechs. Curie.Bio US · Seed fund and drug-hunting accelerator — pairs founding scientists with in-house discovery operators and funds them to a clinical candidate. Lilly Asia Ventures Other · Independent healthcare VC spun out of Eli Lilly's Asia venture arm; the largest dedicated biotech capital pool with heavy China exposure. Beyond Next Ventures Asia · Japan's leading academic-deeptech VC and incubator; funds university spinouts and runs shared wet-lab space for early bio startups. KdT Ventures US · Seed-stage fund with an explicit computational-biology, synthetic-chemistry and applied-physics thesis; invests at company inception. Nucleate US · Non-profit, equity-free accelerator run by and for graduate students and postdocs — the top of the funnel that turns academic biology into companies before any VC is involved. SOSV (IndieBio) US · Multi-stage deep-tech accelerator-VC; its IndieBio programme gives pre-seed biotech founders lab space plus a first cheque, now folded into SOSV NY and SOSV SF. + Gap Nobody in this segment supplies compute or model access as part of the cheque Every non-cash support described in these 14 rows is wet: Beyond Next runs Beyond BioLAB TOKYO shared wet lab, SOSV/IndieBio gives lab space plus $250K for 8%, Curie.Bio pairs founders with in-house drug-hunting operators, Flagship supplies venture-creation staff. No row offers GPU credits, a training cluster, or foundation-model access as an investor-side service -- even the funds whose thesis is computational (Dimension, DCVC Bio, KdT). Caveat: the underlying capability is in the radar, but as portfolio companies in other segments (EvolutionaryScale and Chai Discovery under Foundation models, Lila Sciences under Autonomous labs), not as something an investor here provides; and these rows describe fund closes and headline programme terms, so an informal compute perk could exist without appearing. Absence is from this radar's snapshot, not from the market. + Gap No non-dilutive or asset-backed financing instrument appears anywhere in this segment All 14 rows are equity vehicles (or Nucleate's equity-free grant-style accelerator) -- no venture debt, no milestone-based tranching, no royalty or IP-backed structure, and specifically nothing that finances a biological data asset against its own future licensing revenue. That matters because the radar treats such assets as a distinct business: Proprietary biological data assets for model training holds A-Alpha Bio, Variant Bio, deCODE genetics, Nightingale Health, MedGenome and Protai. So the assets are in this radar and the capital structures for them are not. Caveat: fund announcements rarely enumerate secondary instruments, so this is absence from the snapshot's descriptions rather than proof no such vehicle exists. + Gap No investor here describes a capability for verifying whether a portfolio company's models actually work The operator muscle these firms advertise is biological -- Curie.Bio funds founders to a clinical candidate with in-house discovery operators, Flagship builds companies in-house, Nucleate and SOSV run programmes -- and none of the 14 rows names computational or model-level diligence as a service. Independent validation does exist in this radar, but on the other side of the table: Closed-loop wet-lab validation as a service (Adaptyv Bio, LabGenius, Arctoris, Tierra Biosciences, Nuclera) sells it to companies, not to their investors. Caveat: diligence process is rarely publicised, so this is a gap in what these rows say about themselves, in this radar only. + Gap Asia outside China, and every emerging market, has no dedicated capital pool in this segment -- and the one China-exposed pool is shrinking. HQs here are 8 US, 1 France, 1 Denmark, 1 Japan and 1 China; Lilly Asia Ventures is the only Asia-dedicated fund and its Fund VII closed at $700M against $1.35B in 2021, while Beyond Next Ventures is Japan-domestic at ~$170M. No India, Southeast Asia, Latin America, Middle East or Africa investor appears, despite the radar carrying companies from those geographies elsewhere (MedGenome under Proprietary data assets). Caveat: this may reflect the radar's own selection bias as much as the market, and Nucleate operates across 23 regions in 9 countries without being capital -- so read this as no dedicated pool listed here, not as none existing. + Gap No crossover or public-market vehicle sits in this segment, even though the exits it depends on are public Every row is private venture, evergreen or accelerator capital; the closest thing to a public-market participant is Novo Holdings, and it appears here as the victim of that exposure rather than an investor in it -- its DKK 694B AUM fell 34% year on year because Novo Nordisk's market value fell. Meanwhile this radar contains listed AI-native names whose prices are the actual mark on this thesis (Recursion Pharmaceuticals, Absci and Lantern Pharma under Therapeutics -- in the clinic; Schrodinger under Computational chemistry; Ginkgo Bioworks under Foundries; Moderna under Programmable medicine; Simulations Plus under Trial design). Caveat: crossover and hedge-fund capital is plausibly outside this radar's stated scope, so this is an absence from the snapshot and not a claim that no such fund exists. Point-in-time snapshots, not market censuses. Absence from a radar is not evidence a company does not exist, and funding and traction age fast — re-verify a row against its source before it informs a decision. Not investment, commercial or legal advice.
