AI-Native Drug Discovery: Separating Platform Companies from Software Dressed Up as Biotech

The Headline Numbers

Isomorphic Labs, the Alphabet-spun-out company built around AlphaFold, released a protein-structure-prediction system in early 2026 that reportedly doubles the accuracy of AlphaFold 3 — itself already Nobel Prize-winning work, following Demis Hassabis and John Jumper’s 2024 award for the original model. More than two hundred AI-discovered drug candidates are now somewhere in human clinical trials across the industry. Between fifteen and twenty of them are expected to enter pivotal Phase 3 programs before the end of this year. And the number that should anchor every conversation about this category is this: as of the most recent tally, zero of them have reached FDA approval. Not one.

That gap — between finding a promising molecule and actually curing a disease in an approved, prescribable product — is still enormous, and it is the discipline every investor in this space needs to hold onto even as the headlines get louder. It is worth being explicit about why that gap persists despite the genuinely impressive computational progress: a drug candidate identified or optimized by an AI model still has to survive the same multi-year, multi-phase clinical gauntlet as any conventionally discovered molecule. AI can compress the discovery phase dramatically. It cannot, at least not yet, compress the biology of how long a human body takes to reveal whether a drug is safe and effective, which is where the real time and cost in drug development has always lived.

What the Technology Is Actually Trying to Compress

It helps to remember what the technology is actually trying to compress. A widely cited Tufts Center estimate puts the fully loaded cost of bringing a single drug to market, including the cost of capital and the cost of every failed candidate along the way, at $2.6 billion. Estimates using only publicly disclosed trial costs run lower, in the $985 million to $1.34 billion range, but either way the number is the whole reason the industry has spent the past several years throwing capital at anything claiming to shorten the discovery-to-clinic timeline.

The promise of AI-native drug discovery is not marginal cost reduction. It is a structural rewrite of an industry whose economics have barely moved in three decades, and the specific place that rewrite is supposed to happen is the earliest, most failure-prone stage of the pipeline: target identification and lead optimization, where a traditional discovery program might screen tens of thousands of candidate molecules through slow, expensive wet-lab assays before finding a handful worth advancing. A genuinely capable AI model can, in principle, narrow that search space computationally before a single physical molecule is synthesized, converting a multi-year, multi-million-dollar screening campaign into a matter of months. That is the theory. Whether it holds up at the scale the mega-deals discussed below imply is the actual open question.

Isomorphic Labs: The Category Benchmark

The evidence that something real is happening is genuinely accumulating, even if it has not yet crossed the finish line. Isomorphic Labs raised $600 million in March 2025 in a round led by Thrive Capital specifically to fund its own internal oncology and immunology pipeline into clinical development, alongside continued platform work with pharma partners. The company has been explicit that its internal programs are being advanced with the intention of eventually licensing them out after early-stage trials, rather than building out full commercial infrastructure itself — a capital-efficient model that lets it stay focused on what its own platform is genuinely differentiated at, discovery and early clinical validation, rather than competing with the commercial and manufacturing infrastructure of an established pharmaceutical company.

Demis Hassabis, the company’s chief executive, has been candid about the timeline slipping — he originally projected human trials by the end of 2025, and has since pushed that guidance to the end of 2026. That kind of public schedule slip is, if anything, a point in the company’s favor: it suggests a team being held to genuine regulatory and safety standards rather than one racing to hit a marketing milestone, and it is worth noting explicitly because it is exactly the kind of detail that gets lost in more promotional coverage of the category. A platform that quietly extends its own timeline by a year, rather than rushing an underprepared candidate into the clinic to hit a press-release date, is behaving the way a genuine drug-development organization behaves, not the way a software company under quarterly growth pressure behaves.

Insilico Medicine: The Clearest Proof Point So Far

Insilico Medicine offers the most concrete evidence to date that AI-native discovery can compress real timelines: the company took a drug candidate from initial concept to Phase 1 clinical trial in thirty months, well ahead of typical industry timelines, and has since generated genuine Phase 2a data for its idiopathic pulmonary fibrosis program — a notoriously difficult disease area with a thin existing pipeline of effective therapies, which makes a genuine positive signal there more clinically meaningful than a comparable result in a more crowded therapeutic area would be.

That is the kind of result that moves the conversation from benchmark slide to clinical evidence, and it is not a coincidence that Insilico is also the company at the center of the largest AI-pharma deal ever signed: a collaboration with Eli Lilly worth up to $2.75 billion, announced in March 2026. In the same window, Lilly separately committed up to $1.75 billion to Isomorphic Labs, Novartis signed with Generate:Biomedicines, and Roche committed up to $1.05 billion to Dyno Therapeutics for AI-designed gene therapy vectors, a genuinely distinct application of the same underlying computational approach applied to vector engineering rather than small-molecule or antibody design.

What the Mega-Deals Actually Signal

These are not exploratory pilot programs. When pharmaceutical companies that spend five to six billion dollars a year on R&D start writing checks of that size to AI-native platforms, they are reorganizing how they discover drugs, not experimenting at the margins. It is worth understanding what these deal structures typically look like, because the headline dollar figures are almost always milestone-laden totals rather than upfront cash, and the distinction matters for evaluating how much genuine conviction sits behind the number. A $2.75 billion collaboration typically breaks down into a modest upfront payment, often in the low hundreds of millions, with the bulk of the value contingent on the AI platform actually delivering clinical and regulatory milestones across multiple programs over many years. That structure is not a criticism — it is the standard way large pharma de-risks early-stage collaborations of any kind — but it does mean the headline figures overstate near-term cash commitment relative to how they are usually reported in press coverage of the category.

The Honest Assessment

And yet the honest assessment, echoed across the trade press covering this space through late 2025 and into 2026, is that the field remains in a proof-of-concept phase rather than having delivered a proven paradigm shift. That is not a criticism — proof of concept is genuine progress, and Insilico’s Phase 2a data, Recursion’s advancing FAP program, and Isomorphic’s accelerating benchmark performance are all real, verifiable steps. It is a discipline reminder for anyone writing checks in this category: reward platforms with actual clinical data behind them, not benchmark performance alone, however impressive the underlying model.

Biotech venture funding overall told its own version of this story through early 2026: roughly a quarter of all major US venture rounds tracked in the opening weeks of the year were biotech deals, several of them well above $100 million, reversing what had looked like a multi-year drift of capital toward generative AI infrastructure instead. AI-native drug discovery did not cannibalize biotech funding the way some feared — it became one of the more credible reasons biotech funding held up, precisely because it gave traditional biotech investors a genuine reason to believe the category’s historically poor capital efficiency might finally be improving.

The Underwriting Checklist

That discipline translates into a fairly specific underwriting checklist, one we apply consistently across every AI-drug-discovery term sheet regardless of how compelling the platform demonstration looks in a partner meeting. The platforms worth backing tend to share four traits. They maintain a genuine wet-lab feedback loop — the AI model is trained on and validated against the company’s own experimental data, not purely in-silico prediction divorced from a bench, which matters because a model trained only on public data tends to plateau quickly against problems the public data does not represent well. They own or co-own intellectual property on the output molecules themselves, not merely a fee-for-service prediction engine licensed to whoever pays for access, since a services business built around AI drug discovery captures a fundamentally smaller share of any resulting drug’s value than a company that owns the asset outright. They can point to at least one named clinical-stage asset with real trial data behind it, not only a pipeline of preclinical candidates, because preclinical success in this field has a historically poor track record of predicting clinical success. And their business model survives even in a scenario where the platform never sells another licensing deal — meaning the internal pipeline, not the software business, is the actual source of enterprise value.

The Catalyst We Are Watching For

The first genuine Phase 3 positive readout from an AI-native platform, whenever it lands, will re-rate the entire category overnight — every comparable company’s valuation multiple will move on that single data point, the way clinical-stage biotech has always traded on binary catalysts. Until that event, the discipline for a concentrated fund is to keep evaluating the biology and the clinical data in front of us, not the elegance of the underlying model architecture. The platforms that will still be standing when that first approval finally lands are the ones already treating themselves as drug companies that happen to use AI, not AI companies that happen to make drugs — a distinction that sounds like semantics until you watch how differently the two types of company actually allocate capital, hire, and prioritize between publishing benchmark results and running the unglamorous, expensive clinical trials that are the only thing that will ever actually prove the thesis.