AI Drug Discovery and Life-Science Platforms: A Deep Dive

Core Conclusions

• AI drug discovery has moved from "research showcase" into a "tiered commercialization" phase. The first revenue to materialize comes not from "how many new drugs AI invented" but from the software, data, simulation, clinical and compliance platforms built around pharma R&D workflows: Schrödinger's computational chemistry software, Certara's biosimulation, Veeva's life-science cloud, IQVIA's clinical and data network, and Tempus's clinical-molecular data platform have all already generated verifiable revenue. The FDA has also disclosed that it received 500+ regulatory submissions containing AI components between 2016 and 2023, spanning nonclinical, clinical, post-market and manufacturing stages.

• The links with the clearest real commercial value fall into five categories: First, computational chemistry / structure and molecular design software; second, biosimulation and PK/PD modeling; third, clinical trial design, patient recruitment, and RWD/RWE analytics; fourth, life-science R&D data platforms, ELN/LIMS, knowledge graphs and document-compliance automation; fifth, lab automation and closed-loop optimization. By contrast, assets with "only a foundation model, no proprietary data and no wet-lab loop" are far easier to commoditize.

• AI can indeed shorten early R&D cycles and reduce wasted experiments, but a lift to overall clinical success rates has not yet been fully proven by large-sample industry data. Public cases led by Insilico show that target identification to preclinical candidate nomination can be compressed to roughly 18 months, with screening completed through about 80 molecules; Iambic's published early-model results claim a meaningful improvement in early developability prediction accuracy. What is still genuinely missing: whether these early efficiency gains can be reliably transmitted to Phase II/III and to approval success.

• The profit pool sits mainly with the "pick-and-shovel" vendors and "workflow platforms" in the near term, including software, data, simulation, clinical and R&D-IT infrastructure; the largest mid-to-long-term upside may go to AI-native biotechs that command platform + pipeline + closed-loop experimentation capabilities. Put differently, what is most visible today is ARR, subscription fees, service fees and partnership upfronts; the most nonlinear future value lies in milestones, royalties and proprietary-pipeline NPV.

• "A large headline deal value" does not equal "high revenue quality." For example, Recursion's total potential milestones with Sanofi can reach roughly $5.2 billion, and its collaborations with Roche/Genentech also include large milestone and royalty terms; but what has actually landed is the early upfront plus triggered milestones, not the headline deal value. To judge deal quality, look at: the upfront ratio, the density of stage milestones, who bears clinical costs, the retained royalty level, and whether retained proprietary equity is allowed.

• The companies that truly sit at the platform core are not the "most accurate model companies" but those that simultaneously command proprietary data, an experimental loop, pharma-workflow interfaces, customer relationships, and regulatory context. Representative examples include: Recursion (high-dimensional phenotyping + CRISPR phenomap), Tempus (clinical-molecular data network), IQVIA (global RWD and CRO network), Veeva (life-science industry cloud), Certara (biosimulation with extremely high regulatory acceptance), and Benchling/Dotmatics (R&D data / experiment-record infrastructure).

• Among AI-native biotechs, the ones most worth studying are not those "best at telling a model story" but those that meet four conditions at once: repeat collaborations with pharma; proprietary or partnered programs in the clinic; data assets that can be continuously reused; and the organizational capability to turn model outputs into experimental/clinical decisions. The companies that currently fit this screen most closely include Isomorphic Labs, Insilico Medicine, Iambic, Generate:Biomedicines, and Recursion.

• The excess profit pool of "protein structure prediction" itself is not large; the real profit pool sits downstream. AlphaFold and its ecosystem have dramatically reduced the scarcity of structural information; RoseTTAFold All-Atom has also advanced complex modeling and design. As a result, companies that do pure "structure prediction" have a weakening moat; more valuable are companies that integrate structure, sequence, function, manufacturability and wet-lab feedback into a production system.

• AI protein design, antibody design and molecule generation are among the corners most prone to "bubbles", because demos are easy while real pharmaceutical developability, CMC, toxicology and clinical translation are hard; by contrast, R&D data platforms, biosimulation, clinical-trial AI, compliance-document automation and R&D operating systems are more likely to compound over the long run.

• Among public companies, the AI beneficiaries with the strongest revenue validation are mostly not pure AI drug-discovery companies. From the angle of "revenue already landed, customers already validated, gross margin visible," Veeva, IQVIA, Certara, Schrödinger, Tempus, Thermo Fisher, Danaher, 10x Genomics and Illumina are more trackable than most preclinical AI-native biotechs.

• The classic risk profile of "strong AI narrative, weak commercialization" includes: revenue extremely dependent on one-off deal recognition; undisclosed platform usage frequency and repeat-purchase rates; no recurring large-pharma customers; no proprietary/partnered programs in the clinic; cash burn outpacing deal realization; valuation that pre-spends clinical success rates. For such companies, investors most need to guard against "the milestone illusion" and "the model-capability illusion." This risk is especially pronounced in some early AI-native biotechs, automated labs, and "general-purpose bio foundation model" companies.

• M&A is already validating the value of the "life-science R&D software layer," not just the "model layer." Siemens acquired Dotmatics at an enterprise value of $5.1 billion, explicitly to expand its AI-driven life-science R&D software footprint; public reporting also indicates that Roche intends to acquire PathAI for up to $1.05 billion, which, if completed, would further validate the strategic value of AI pathology / clinical R&D software assets.

• The biggest catalysts over the next 12–24 months are not papers but three types of events: the continuity of AI-discovered molecules in Phase II readouts; large-pharma collaborations upgrading from "exploratory partnerships" to "transfer of development and commercialization rights"; and larger M&A among life-science R&D-IT and automation platforms. Representative trackable events today include: Insilico's clinical progress, AbCellera's ABCL635/575 readouts, Generate's GB-0895 Phase III progression, Isomorphic's first entry into human studies, and more deals resembling Dotmatics/PathAI.

• The biggest long-term risk is not that models are too small but that biology is too complex, that the translation step carries too much noise, and that technical barriers leak out once foundation models are rapidly open-sourced. The FDA/EMA are currently both emphasizing risk-based approaches, interpretability, data governance and applicability boundaries—which essentially says that regulators accept AI but do not accept "a black box replacing validation."

• The point most easily misunderstood by the market is: "AI drug discovery" is not a single track but a composite of software-platform value, data-asset value, experimental-loop value, clinical-validation value, and drug-pipeline value; companies that nominally all go by "AI biotech" can have completely different economics.

Industry-Chain Landscape and Profit-Pool Judgment

Validation gradient precedes valuation. In this theme, what matters most is not "the AI is strong" but which validation tier a company sits at: Scientific breakthrough (e.g., AlphaFold 3, RoseTTAFold All-Atom) → platform partnership (e.g., Isomorphic, Recursion, Generate, Insilico signing with large pharma) → revenue landing (e.g., Schrödinger, Certara, Veeva, IQVIA, Tempus, BenchSci) → clinical validation (e.g., Insilico, AbCellera, Generate, Relay) → drug approval (among the samples covered here, there is not yet a clear case of a pure AI-native biotech using "AI discovery" as its core narrative being formally approved for market by a mainstream regulator; this point still needs ongoing validation).

The core judgment on profit-pool distribution is: the most profitable layer in the near term is "digitalizing and simulating the R&D workflow"; the mid term favors "platform + service + partnership"; the long term favors "platform + pipeline + equity." The pure-model layer is the easiest to be diluted by open source and cloud vendors; the pure-CRO-labor layer is the easiest to face price pressure from efficiency gains; the true long-term high barrier lies in proprietary data × lab automation × pharma-workflow embedding × clinical/regulatory usability. This is exactly why Siemens acquired Dotmatics and pharma buys Veeva/Benchling/Certara/IQVIA, rather than only buying a general-purpose large model.

Simplified judgment: Highest near-term certainty belongs to "workflow-layer" companies like Veeva, IQVIA, Certara, Schrödinger, Tempus, Dotmatics/Benchling; largest mid-term optionality belongs to Isomorphic, Insilico, Generate, Iambic, Recursion, AbCellera; indirect beneficiaries that look more like pick-and-shovel vendors are Thermo Fisher, Danaher, 10x Genomics, Illumina, Opentrons; most exposed to disruption are the traditional discovery-outsourcing and manual processes with low technology density, low data content and low automation penetration.

Business Models, Technology Roadmaps and Valuation Frameworks

How AI drug-discovery companies make money. The revenue models the industry has already validated can be grouped into six categories: software subscription / SaaS (Schrödinger, Veeva, Benchling, Dotmatics, Certara); platform licensing and data-access fees (Tempus, IQVIA, and parts of the BenchSci/Tempus/Recursion model); collaborative R&D revenue (Recursion, Generate, Insilico, AbCellera); milestones and royalties (Recursion, AbCellera, Insilico, Generate, Isomorphic); proprietary-pipeline value (Insilico, Generate, AbCellera, Recursion, Relay); M&A exit value (Dotmatics acquired by Siemens; PathAI reportedly acquired by Roche for up to $1.05 billion).

Which model is better. Looking only at revenue quality, the ranking is usually: subscription software > deeply embedded platform fees > repeatable analytics services > upfront-based partnerships > milestones/royalties > single proprietary-pipeline valuation. But if you look at upside, the order almost reverses. The optimal mix is usually neither pure software nor pure pipeline, but software/platform cash flow + some high-quality pipeline equity. Companies like Schrödinger, Certara, Tempus and Veeva are more readily accepted by the secondary market precisely because at least part of their revenue is predictable; pure AI-native biotechs, by contrast, must trade higher volatility for higher upside.

The advantages of the platform + pipeline hybrid model are: first, it lets pharma pay for the platform, easing cash burn; second, it lets platform output stay on the company's own books as a proprietary pipeline, capturing nonlinear returns; third, it more quickly validates whether the platform is repeatable. The risks are: high organizational complexity, heavy capex, volatile revenue recognition, and the tendency for a platform company to gradually "degrade" into an ordinary biotech. Recursion, AbCellera, Generate and Insilico are all going through this tension.

How to judge collaboration-agreement quality. A high-quality agreement usually features: a higher upfront, clearer near-term technical/development milestones, retained royalties or co-development equity, the large pharma bearing subsequent clinical and commercialization costs, and customers showing repeat-purchase / expansion behavior. For example, Isomorphic's collaborations with Lilly/Novartis carry total potential value approaching $3 billion; Generate's multi-target collaboration with Novartis carries total potential value over $1 billion, with an upfront of $65 million; Insilico's new global collaboration with Lilly includes a $115 million upfront and up to $2.75 billion in total potential value; Recursion has already received $134 million in upfront plus progress milestones from Sanofi, and $213 million in upfront and milestones from Roche/Genentech.

The value tiering of technology roadmaps can be summarized as follows:

Three Scenario Forecasts

Conclusions on the valuation framework:

• Software-platform type: prioritize ARR/subscription-revenue growth, gross margin, net retention, customer count, and depth of industry embedding.

• Platform + partnership type: look at upfront quality, the density of triggerable milestones, partnership repeat rate, and the pace of proprietary-program advancement.

• Pipeline AI-biotech: ultimately still reverts to traditional biotech rNPV, except that whether the platform can raise success rates affects the success-probability assumption.

• Data-asset type: look at whether the data is exclusive, whether it is structured, and whether it can be used directly for pharma decisions and regulatory acceptance.

• What should least be valued on its own is "the foundation model itself," because structure prediction and general-purpose generative models are being open-sourced and commoditized ever faster.

A Deep Breakdown of Sub-Tracks

The table below compresses each track along "track logic—revenue conversion—barriers—risks—investment attractiveness." Scores reflect this report's research judgment, out of 10.

The tracks most prone to bubbles: general-purpose bio foundation models, "pure generation" companies without a wet-lab loop, single-point structure prediction, and AI-native biotechs that sell only vision without disclosing customer repeat-purchase and clinical progress.

The tracks most likely to produce real compounding: life-science R&D software platforms, clinical and real-world data, biosimulation, workflow-level AI, lab automation with a data loop, and protein/antibody-design companies that genuinely enter the clinic and retain equity.

Master List of Investment Targets and Company Tiering

Master List of Selected Public and Important Private Targets

Note: valuation metrics preferentially use market cap/share price as of 2026-05-19 plus disclosed full-year 2025 revenue for a rough basis; where complete current-period figures for EV/EBITDA, cash burn or R&D expense are not retrieved in this report, they are marked "needs further validation." The "benefit path" and "disruption path" are this report's research judgment.

Company Tiering and Investment Priority

Tier A: Core direct beneficiaries of AI drug discovery Veeva, IQVIA, Certara, Schrödinger, Tempus, Benchling, Dotmatics/Siemens. Rationale: they have either formed clear subscription/platform revenue, or command high-value clinical/R&D data networks, or are already embedded in pharma's core workflows.

Tier B: Clear beneficiaries, but with valuation, clinical or commercialization risk Recursion, AbCellera, Generate, Insilico, Iambic, 10x Genomics, WuXi AppTec. Rationale: strong platforms, but either high cash burn, insufficient clinical validation, or still-volatile revenue.

Tier C: AI mainly used for efficiency; weak near-term financial upside Thermo Fisher, Danaher, Illumina, Tigermed, BGI Genomics. Rationale: AI is an enhancer rather than a primary revenue source; benefits are more indirect.

Tier D: Strong narrative but still insufficient evidence of real benefit Some general-purpose foundation-model companies, pure target-discovery companies, and early AI-biotechs with no repeat pharma collaborations / no clinical advancement / no revenue disclosure. Rationale: model capability has not yet translated into stable revenue or clinical value.

Tier E: Traditional links that may be compressed by AI and automation Low-automation, low-data-accumulation, low-technology-content manual discovery outsourcing and documentation/translation/monitoring processes. Rationale: what AI and automation replace first are standardized, repetitive, low-differentiation processes.

Scoring Model and Key-Company Ranking

Positive scoring model

• Direct exposure to AI drug-discovery revenue or pipeline value: 20%

• Barriers across data, models and the experimental loop: 20%

• Pharma-collaboration and customer quality: 15%

• Pipeline quality and clinical validation: 15%

• Financial quality and cash runway: 10%

• Market size and growth upside: 10%

• Valuation reasonableness: 10%

Reverse-risk model

• Insufficient clinical validation: 25%

• Insufficient revenue durability: 20%

• Cash burn and financing risk: 20%

• Platform effectiveness not yet proven: 15%

• Risk of pharma building in-house or open-source substitution: 10%

• Overvaluation: 10%

Deep Analysis of Key Public Companies

To control length, the following is compressed by "the public samples most worth further study," focusing on direct exposure, platform barriers, revenue validation, valuation and risk.

Risks, Expectation Gaps and Final Conclusions

Which companies have already fully priced in AI expectations. From the current market narrative and valuation structure, Recursion, AbCellera, some high-heat AI-native biotechs, and general-model companies that still lack stable revenue are more prone to "premature capitalization of platform value"; meanwhile, if future clinical delivery falls short, valuations will be more fragile. Conversely, Certara, IQVIA, and some life-science R&D software assets have not fully reflected in the secondary market AI's potential as an efficiency and category-expansion engine.

Which companies may still have expectation gaps.

• Certara: the main reason it is most underrated is that the market still sees it as "professional software + services" rather than one of the decision platforms with the highest AI + regulatory acceptance in drug discovery.

• IQVIA: its combination of RWD + CRO + AI is closer to large pharma's budget center than most "single-point AI companies."

• Schrödinger: the market is more sensitive to its pipeline volatility and easily underrates the industry-gateway nature of its software platform.

• Veeva: outsiders easily see it as traditional SaaS, but the embedding depth of life-science-industry-specific AI agents may let it keep winning incremental budget.

• 10x Genomics: if spatial/single-cell data becomes the "new base" for AI target discovery and patient stratification, its data-infrastructure value will be re-rated.

Which traditional companies are most likely to benefit. Thermo Fisher, Danaher, Illumina, 10x, IQVIA, Veeva, WuXi AppTec, Tigermed. These companies benefit not because "they will build the best models themselves" but because AI will increase demand for high-quality data, automation, compliance workflows, clinical execution networks, omics infrastructure and R&D software.

Which traditional companies may be disrupted. The most likely to face pressure are: first, low-tech services that rely on large amounts of manual screening, manual documentation and manual monitoring; second, discovery outsourcing without proprietary data and software control; third, low-end experimental processes lacking automation capability. AI does not have to replace all CROs to compress their profit pools; merely raising R&D efficiency inside pharma and at leading platforms will change pricing power.

Systemic risk checklist.

• Commercialization below expectations: customers treat AI as a pilot rather than a core procurement.

• Clinical failure: the platform improves early efficiency but does not improve human translation.

• Platform effectiveness cannot be proven: only individual cases can be told, not replicated at scale.

• Insufficient revenue durability: partnership revenue is highly dependent on milestones.

• Cash burn and refinancing risk: especially for the platform + pipeline hybrid model.

• Risk of large pharma building in-house and open-source substitution: especially the structure-prediction and general-model layers.

• Data bias and data-sovereignty risk: constraints on RWD, EHR and cross-border data use.

• Lab automation is capex-heavy and slow to deploy.

• Regulatory uncertainty: AI entering organization-level GxP and filing still needs more standardization.

Final Conclusions

The importance of AI drug discovery and life-science platforms within the AI industry chain lies not in how many "star drugs" it will produce first, but in its being one of the few scenarios that can genuinely embed AI into high-margin, heavily regulated, long-cycle, decision-intensive industry workflows. It is slower than most general-purpose AI applications, but once it enters core workflows, stickiness is stronger and switching costs are higher.

The five sub-tracks most worth watching

• Life-science R&D software and industry cloud

• Clinical data/RWD/patient-recruitment platforms

• Biosimulation and computational chemistry

• Protein/antibody-design platforms with a wet-lab loop

• Automated labs and R&D data infrastructure.

The ten public companies most worth deep study Veeva, IQVIA, Certara, Schrödinger, Tempus, Recursion, Thermo Fisher, Danaher, 10x Genomics, WuXi AppTec.

The ten private/primary-market companies most worth tracking Isomorphic Labs, Iambic, Xaira, Cradle, Benchling, BenchSci, BigHat, PathAI, EvolutionaryScale, Profluent.

The five points most easily misunderstood by the market

• An AlphaFold/foundation-model breakthrough does not mean the profit pool will stay at the model layer.

• AI shortening early discovery time does not mean it will equally shorten approval and time to market.

• A several-billion-dollar collaboration headline does not equal real revenue.

• A "platform company" without recurring revenue and customer embedding may, in the end, be just a high-valuation biotech.

• Data assets and the experimental loop are usually more important than the model itself.

The metrics most worth tracking over the next 6–12 months

• Phase II / pivotal PoC data from AI-native biotechs

• Newly triggered near-term milestones and add-on deals within large-pharma collaborations

• ARR, software-revenue growth, customer retention and large-customer expansion at software platforms

• Large M&A among R&D data platforms / automation platforms

• Cash balance, net cash burn, and implied runway.

Platform-type winners Veeva, IQVIA, Certara, Schrödinger, Benchling, Dotmatics/Siemens, Tempus. AI-native biotech challengers Isomorphic, Insilico, Recursion, Generate, Iambic, Xaira, AbCellera. Pick-and-shovel vendors Thermo Fisher, Danaher, 10x, Illumina, Opentrons, WuXi AppTec. Pseudo-beneficiary or high-risk profiles Companies that only showcase model capability, with no repeat customers / no clinical advancement / no data loop / no disclosure of revenue quality.

Open Questions and Limitations This report has tried to prioritize public information as of 2026-05-19, but several items remain not fully disclosed or in need of further validation:

• Most private companies lack verifiable ARR, gross margins and cash burn;

• Some European, Japanese, Korean and Indian companies do not separately disclose AI-drug-research revenue contributions;

• For a narrower and deeper next step, I would suggest prioritizing any one of the following directions:

AI protein design, AI small-molecule discovery, automated labs, AI clinical trials, AI + CRO, AI pharma R&D platforms, multi-omics AI, AI-native biotech valuation.

This report is based on public information and does not constitute investment advice. Markets carry risk; invest with caution.