The short version
The mistake I see when people talk about TechBio is that they treat it like a single market. It is not. TechBio is better understood as a stack. Some companies are building the data layer for biology. Some are building models. Some are building automated labs. Some are building new therapeutic assets. Some are building clinical infrastructure. Some are rebuilding how pharma companies make decisions.
That distinction matters because the success factors are different. A molecular design company does not win the same way a lab automation company wins. A clinical trial optimization platform does not have the same evidence burden as a protein design platform. A pure software platform does not monetize like a platform plus pipeline biotech. But underneath all of these differences, most serious TechBio companies need the same foundation: proprietary data, biology-aware models, experimental validation, workflow integration, translational credibility, and a business model that survives long biological timelines.
My main thesis is simple: the best TechBio companies will not just be AI model companies. They will be learning systems. They will turn experiments into better data, data into better models, models into better experiments, and experiments into stronger biological products.
Why this market is hard to map
Drug discovery is one of the strangest economic games in technology. A company can spend years and billions of dollars before it knows whether the underlying scientific bet worked. Most drugs fail before approval, and many approved drugs still do not earn back their development costs. The upside is enormous when a therapy works, but the path is slow, uncertain, regulated, and biologically unforgiving.
That is the backdrop for AI × TechBio. The promise is not just that AI can generate molecules faster. The deeper promise is that software, automation, and biological data can change the shape of the entire discovery and development process. Instead of a slow, linear path from hypothesis to molecule to preclinical testing to clinical testing, TechBio pushes toward iterative systems where each round of data improves the next round of prediction.
There is also a regulatory tailwind worth naming. In April 2025 the FDA announced a plan, paired with a five-year roadmap, to phase out its animal-testing requirement, beginning with monoclonal antibodies and expanding to other biologics and small molecules over time. In its place the agency will accept New Approach Methodologies (NAMs): AI-based computational toxicity models, organ-on-chip systems, and human-relevant cell and organoid assays, weighed together as evidence. That is a direct tailwind for the drug-safety, virtual-patient, and validation-focused companies in this map. The regulator is, slowly, starting to accept the kind of evidence TechBio companies are best positioned to generate.
This is why I prefer to map the industry by function rather than by buzzword. TechBio includes AI co-scientists, molecular design engines, virtual cells, protein design platforms, drug safety systems, self-driving labs, clinical trial optimization platforms, synthetic biology foundries, diagnostics companies, biosecurity tools, and data infrastructure companies. These companies overlap, but they are not interchangeable. A useful map needs to answer three questions: where does the company sit in the biological workflow, what kind of moat is it building, and how does it capture value?
What every TechBio company needs
Before dividing the market into verticals, it is worth defining the common requirements. In my view, almost every meaningful TechBio company needs some version of the following capabilities.
| Capability | Why it matters |
|---|---|
| Proprietary data | Biology is not like internet text. High-signal biological data is slow, expensive, messy, and often physically generated in a lab. Companies with unique datasets can build a durable advantage. |
| Biology-native models | Many architectures were invented for language or images. Biology requires models that can reason across molecules, cells, proteins, patients, time, and perturbations. |
| Experimental feedback loops | Predictions have to be tested. A real platform should learn from wet lab or clinical feedback rather than remain a static model demo. |
| Translational credibility | The hardest question is not whether a model can optimize an assay. It is whether that assay predicts human biology, patient response, safety, and clinical outcomes. |
| Workflow integration | Scientists will not adopt tools that sit outside their actual research process. The strongest products embed into lab, data, clinical, or R&D workflows. |
| Commercial clarity | A company needs to know whether it is selling software, services, platform access, drug assets, diagnostics, manufacturing capacity, or some hybrid of these. |
| Cross-functional talent | The strongest teams combine ML, biology, chemistry, translational medicine, product, automation, regulatory thinking, and business development. |
A lot of AI biology companies start with a model story. The stronger ones eventually become data generation companies, experiment design companies, translational biology companies, or platform plus pipeline companies. The strongest ones connect all of these pieces into a compounding loop.
The market map
I would organize the market into nineteen major verticals. The boundaries are fuzzy. A virtual cell company may also be an AI drug discovery company; a protein design company may also be a delivery company; a modern lab company may be infrastructure for almost every other category. The point of the map is not perfect taxonomy. The point is to understand where leverage is being created. The full company index for every vertical lives in the appendix; here I summarize the thesis of each.
1. AI co-scientists. Agentic research platforms that move scientists from literature review and hypothesis generation into experiment design, analysis, and execution. The interesting version is not a chatbot. It is an operating layer that remembers prior work, orchestrates tools, and helps scientists decide what to do next.
2. Bio AI infrastructure and scientific operating systems. As more biological foundation models appear, the bottleneck shifts from availability to usability. These companies make models usable, composable, deployable, and connected to real data systems, and make biology data ready for modeling in the first place.
3. Modern labs and self-driving experimentation. The experimental operating layer: protocols, instruments, robotics, scheduling, and data capture turned into programmable workflows. Strategic because every AI biology company eventually needs better experiments and faster iteration.
4. Drug safety and ADMET. One of the clearest places where earlier prediction saves time and capital, using AI, causal reasoning, ADMET prediction, organ-on-chip, and stem-cell systems to catch toxicity and developability issues before expensive clinical testing.
5. Clinical trial optimization. Attacking the cost and execution problems in trials: protocol design, amendment prevention, faster startup, recruitment, site selection, and study management. More candidates do not help if the clinical machine stays slow and expensive.
6. Virtual patients and virtual clinical trials. Simulating how patients or cohorts respond to therapies to improve translation, stratify patients, select endpoints, and test trial designs in silico. Some are purely computational; others pair organoids, tissue models, and robotics with AI.
7. AI small molecule and chemistry platforms. Small molecules remain central: manufacturable, often oral, able to reach intracellular targets. These platforms cover target discovery, molecule generation, lead optimization, binding prediction, and chemistry planning fused with experimental data.
8. Molecular glues and targeted degradation. A different way to drug biology: instead of inhibiting a target, induce or stabilize protein-protein interactions, often routing a disease protein to degradation machinery, opening previously undruggable target classes.
9. RNA therapeutics, gene therapy, and delivery. Limited not only by design but by delivery. Spans RNA-targeted small molecules, RNA structure, guide RNAs, programmable LNPs, BBB-crossing conjugates, intron-enhanced expression, and cell/gene therapy tooling.
10. Protein therapeutics and programmable biology. The shift from optimizing known antibodies to generating new binders, enzymes, cages, ADC logic, and proteins from scratch, combining foundation models, structure prediction, wet-lab validation, and developability screening.
11. Nature-derived discovery. Nature has produced enormous chemical and biological diversity, only a sliver of it explored. These companies mine metagenomics, plant metabolites, microbiome data, and microbial chemistry with AI for new therapeutic starting points.
12. Molecular structure prediction. The computational substrate beneath much of AI drug design, now moving beyond static structures toward complexes, binding affinity, conformational dynamics, intrinsically disordered proteins, and direct design integration.
13. Physics-first approaches. Leading with quantum mechanics, molecular dynamics, and neural network potentials. The argument: first-principles methods generalize where data is sparse and chemistry moves out of distribution.
14. Virtual cells. Simulating how cells respond to perturbations by integrating genes, proteins, metabolism, signaling, imaging, single-cell, spatial, and clinical context. If it works, scientists run part of the experiment computationally before touching the lab.
15. Immunology and immune engineering. The immune system is dynamic, patient-specific, and hard to rule-model. These companies map immune cell states, discover antigens, model tumor evolution, and design immune-targeted therapies and TCR bispecifics.
16. Drug rescue and asset development. A different premise: many promising compounds already exist but were abandoned, mis-matched to patients, or never understood mechanistically. AI helps source, evaluate, redesign, reposition, or develop them more efficiently.
17. Synthetic biology, cell engineering, and biomanufacturing. TechBio is not only pharma. Using AI, automation, genome engineering, enzyme design, and fermentation to make biology engineerable across chemicals, materials, food, agriculture, and therapeutics.
18. Clinical data, diagnostics, and precision medicine. Connecting AI to real patients: multimodal clinical data, pathology, liquid biopsy, stratification, real-world evidence, and trial matching. A different evidence burden, because the buyer may be a physician, hospital, payer, or pharma team.
19. Biosecurity, regulatory, and translational infrastructure. If AI makes biology easier to design, the industry needs stronger systems for safety, validation, regulatory readiness, and governance. Easy to overlook, increasingly important as generative biology gets more powerful and accessible.
The broader ecosystem
A startup map is incomplete without the larger institutions shaping the market. Public companies, mega-unicorns, Big Tech infrastructure builders, and Big Pharma internal AI divisions each play a different role, and I would keep them separate rather than lumping them together.
Public and mega-unicorn platform pioneers created many of the categories younger startups now build within: Recursion (phenotypic screening and massive datasets), Schrödinger (physics-based chemistry), Exscientia and BenevolentAI (early AI discovery), Absci and Relay (biologics and protein dynamics), Ginkgo (synbio platform thesis), and Owkin, Tempus, and ConcertAI (clinical and genomic data as a platform).
Big Tech and foundational AI infrastructure sits underneath much of the market: NVIDIA (BioNeMo turning accelerated computing into biology-specific model infrastructure), Google DeepMind (AlphaFold, with Isomorphic translating it into drug design), Microsoft Research, IBM Research (MoLFormer), EvolutionaryScale (the Meta/FAIR protein-language lineage), and ByteDance and Tencent (proof this is not only a US and Europe story).
Big Pharma internal TechBio engines are no longer just buying tools. They are acquiring AI companies and building internal platforms: BioNTech + InstaDeep, Sanofi, Genentech + Prescient Design, Moderna, Amgen Generative Biology, Roche + Flatiron and Foundation Medicine, Eli Lilly + TuneLab, and the Novartis AI Innovation Lab. This is why startups must be precise about their wedge. Pharma buyers increasingly have internal AI capability of their own.
How TechBio companies capture value
The business model question is one of the hardest parts of TechBio. In traditional biotech, a company creates enormous value through a single drug. In software, a company creates value through recurring revenue and distribution. TechBio is stuck between these worlds: the product can look like software, but the evidence bar often looks like biotech.
Platform plus pipeline
The dominant model. A company builds a reusable platform, uses it to generate or optimize assets, and captures value through partnerships, co-development, licensing, milestones, royalties, or its own products. The appeal: if the platform really can discover better medicines, the largest upside comes from owning some of the assets. The downside is capital intensity: a pipeline means paying for chemistry, biology, toxicology, manufacturing, regulatory work, and clinical trials, and surviving long enough for the molecules to become convincing.
Pure platform
Pure platform companies do not necessarily own a drug pipeline. They sell model access, workflow software, lab automation, data infrastructure, clinical infrastructure, or decision support. The pushback is that there are fewer pharma buyers than generic SaaS customers, and seat-based pricing caps the market. But I think that critique misses how the market may evolve. If biological AI becomes central to R&D, pricing moves beyond seats: usage-based pricing scales with experiments, molecules, protocols, or simulations run; outcome-based pricing lets a platform share in the value it creates when attribution is clear. Hybrid pricing could become the norm for platforms that operate a whole workflow.
Wedge to platform, and beyond pharma
The most realistic route for many companies is wedge to platform: start with a narrow use case where the value is obvious, then expand into adjacent workflows. A clinical trial startup may begin with recruitment and expand into protocol design, site selection, monitoring, and analytics. Pure platform companies can also expand the market itself: agriculture, food and nutrition, industrial biology, diagnostics, consumer health, and materials science. Cradle is a good example of a protein-engineering platform with relevance beyond therapeutics. Biology is becoming a design and manufacturing substrate, not just a drug-discovery substrate.
CRO budgets and the future of outsourced R&D
One of the most interesting questions is whether TechBio can capture wet-lab and clinical-trial budgets, not just software budgets. On the preclinical side, AI makes experiments more targeted. On the clinical side, AI can fit into functional service partnerships, data management, monitoring, trial design, and patient selection. I do not think traditional CROs disappear quickly. Relationships are entrenched, workflows are regulated, and change management is slow. But AI-native tools can enter through specific functions and expand over time.
The technical stack underneath the market
The company map is useful, but the deeper question is what technical capabilities separate strong TechBio companies from weak ones. I would organize that around data, algorithms, infrastructure, workflows, and validation.
Data: the real bottleneck
The data problem in biology is fundamentally different from consumer AI. Biological data often has to be generated physically. You cannot scrape a complete representation of human biology from the internet. You need assays, samples, instruments, longitudinal follow-up, protocols, metadata, and careful experimental context. Some biological processes simply take time to observe. A strong data moat has six attributes: scale and diversity, velocity, multimodality and richness, quality, novelty of biology or technique, and cost-effective generation.
Algorithms: biology needs generalization
The algorithmic challenge is generalization. Biology is sparse, discontinuous, context-dependent, and dynamic, and the model must work where training data is incomplete. That is why biology-native architectures matter. CellType adapts structured biological data into language-like formats (Cell2Sentence); BioState AI and EchoJEPA push toward architectures (JEPA-style) that learn underlying biological relationships rather than memorizing noisy measurements, with scGPT a useful contrast, since gene expression is not simply language. Cost matters too: Boltz focuses on low-cost inference so more designs can be evaluated; Converge Bio and Latent Labs work on model economics and fast inference. Multi-model systems matter because different models win different tasks: Tangram's LLibra selects and updates models as performance changes. The hardest layer may be translation: linking biological context to patient outcomes and clinical response, and moving target discovery beyond crowding by starting from data (Scripta begins from disease-associated transcriptional signatures rather than predefined targets).
Infrastructure and workflows: the lab in the loop
I do not think AI eliminates the wet lab. I think it changes what the wet lab is for. Instead of brute-force screens, the lab becomes part of a feedback loop: models propose, experiments test, the data improves the model, and the next experiment gets smarter. The best lab-in-the-loop systems optimize for speed, quality, context capture, and curiosity: Onava on rapid iteration, Synteny and Noetik on cheap-but- powerful high-throughput data, Relation Therapeutics on co-locating AI and wet-lab teams around translational fidelity, and Phylo on capturing the tacit reasoning behind experiments. The most exciting version is not just automation but curiosity-driven automation: systems that explore high-information experiments and surface features humans would miss.
Validation: proving the platform before the pipeline matures
Early validation is a defining challenge. Clinical readouts are the clearest proof, but early-stage startups rarely have them, so they need other ways to show the platform is not just fitting noise. I think about it in three buckets: standardized preclinical packages (consistent assays around potency, specificity, safety, and developability, benchmarked over time, with Synteny disciplined here); human validation outside formal trials (Outpost Bio using human microbiome interventions in consumer-health contexts to validate faster than a pharma trial allows); and predicting unseen data (if a platform predicts an insight a pharma company already knows internally but has not published, that is powerful. CellType's work around Senhwa Biosciences and Orakl's trial-outcome predictions illustrate the logic). The practical lesson: the evidence bar depends on the biology. Known biology with reliable translation models is easier to validate; novel biology with huge unmet need can justify early risk; the dangerous zone is the middle, where existing drugs raise the bar, translation models are unreliable, and the upside does not justify the uncertainty.
My synthesis: the winners will own a loop, not just a layer
After looking across these verticals, my biggest takeaway is that the next generation of TechBio winners will not be defined by a single model architecture or a single asset. They will be defined by the loop they control. A small molecule platform that only generates molecules is less interesting than one that connects generation, synthesis, assay results, ADMET, and clinical translation. A protein design platform that only outputs sequences is less interesting than one that connects expression, stability, binding, immunogenicity, developability, and manufacturing. A lab automation company that only runs protocols is less interesting than one that turns every protocol into reusable, structured, model-improving data.
The most important phrase in TechBio is closed loop. The company that closes the loop between data, models, experiments, and decisions gets compounding returns. Every experiment makes the dataset better. Every dataset makes the model better. Every model makes the next experiment smarter. Every smarter experiment creates stronger evidence. That is the flywheel.
This also explains why TechBio is hard. It is not enough to hire ML engineers and apply foundation models to public datasets. It is not enough to build a nice wet lab. It is not enough to have a platform story. The hard part is orchestrating all the layers at once: data generation, model development, scientific reasoning, experimental execution, translational validation, and commercial value capture. That is also what makes the category so interesting. TechBio is not simply AI for biotech. It is the attempt to make biology more programmable, measurable, and engineerable. The companies that matter most will be the ones that turn biology into a compounding learning system, making every experiment, dataset, workflow, and clinical insight improve the next one.
Why function-first, and where I'd put my chips
Most market maps I've seen slice the space by modality (small molecule vs. protein vs. RNA), by pipeline stage (discovery, preclinical, clinical), or by company type (startup vs. incumbent). That's a reasonable way to organize a directory, and it's how most BD teams already think about the space. But it groups companies that have almost nothing in common just because they share a modality label, and it separates companies that are solving the exact same problem in different biological contexts. An RNA-delivery startup and a protein-design startup can have nearly identical data, model, and validation challenges; two small-molecule companies in the same "modality" bucket can have completely different sources of defensibility. Slicing by function, by what layer of the stack a company actually builds, groups companies by what they need to get right to win, which is the question I actually care about.
I think the AI co-scientist (literature-mining layer) can become an overrated area. Synthesizing papers and surfacing hypotheses is useful, but it's an interface on top of public knowledge, not a closed loop. Nothing about the product gets structurally better just because more people use it, and general-purpose model providers can absorb most of this capability as a feature rather than a company. If I had to bet on one part of the stack being underpriced relative to its importance, it's the data layer: companies generating novel, proprietary, multimodal biological data at scale, like Tahoe and Basecamp. Everyone above them in the stack is downstream of what they produce, and unlike a model or a pipeline asset, a good data-generation engine keeps compounding in value the longer it runs. That's the layer I'd expect to look most different, and most valuable, five years from now.
Appendix: company index by vertical
This index keeps the essay readable while making sure no company gets lost. A few names appear in two verticals on purpose, because the company genuinely spans both: Xaira does protein and small-molecule design and builds a virtual-cell model, and Causaly is both a literature-mining co-scientist and a piece of research infrastructure. One apparent duplicate is actually two unrelated companies that share a name — Nabla Bio (de novo antibody design) and Nabla (the clinical-documentation assistant) — listed separately below.