200 Drugs, Zero Approvals: AI Biotech's Reckoning Is Here

The number circulating through biotech boardrooms this spring is 200. That's how many AI-discovered drug candidates are currently enrolled in clinical trials — 94 in Phase 1, 56 in Phase 2, 15 pressing into Phase 3 — and zero across the FDA finish line. The pipeline has never been larger. The approval column has never been emptier. Both facts matter, and the industry is increasingly unwilling to pretend otherwise.

This is where AI biotech stands heading into the second half of 2026: past peak hype, deep into the "builder" phase, confronting the basic biological truth that molecules take years to kill or cure people and no amount of compute changes that.

Robin, AlphaFold, and the New Discovery Stack

The clearest signal that AI drug discovery has genuinely shifted — not just in pitch decks but in the lab — came in May, when a Nature-published study introduced Robin, described as the first AI system to run a fully autonomous discovery-and-validation loop using real wet-lab experiments as feedback. Applied to dry age-related macular degeneration, Robin surfaced ripasudil, a glaucoma drug, as a repurposing candidate and confirmed efficacy in vitro without human-directed hypothesis generation at any step.

That's a meaningful architectural leap. Earlier AI discovery systems handed researchers a ranked list of candidates and stepped back. Robin runs the loop — propose, test, revise — treating laboratory outputs as training signal in real time. It's the distinction between a very smart search engine and something closer to an autonomous research agent.

The underlying infrastructure making this possible is AlphaFold 3, released by Google DeepMind in late 2024 and now embedded in the stack at Isomorphic Labs and dozens of smaller shops. Where earlier versions predicted protein folding, AF3 models the binding geometry between small molecules and their targets — the exact prediction that historically required years of crystallography and a fair amount of luck. Isomorphic hasn't advanced compounds into human trials yet, but their leverage is structural: if you can reliably predict how a molecule binds before synthesizing it, you collapse the most expensive part of medicinal chemistry.

The Companies with Something to Prove

Three names define the clinical frontier right now.

Insilico Medicine holds the most credible end-to-end claim. Its compound for idiopathic pulmonary fibrosis — a disease with few good options and a median survival of three to five years post-diagnosis — went from target identification to Phase II in under 30 months. That timeline would have been considered fantasy a decade ago. Rentosertib, a second candidate, is advancing in parallel. Insilico is the company most likely to give the sector its first true proof point.

Recursion Pharmaceuticals, after its merger with Exscientia, is the largest AI-native biotech by headcount and ambition. It's also the one nursing the most public wound: REC-994, its lead candidate for cerebral cavernous malformation, was discontinued last year after long-term data failed to confirm earlier efficacy trends. The setback didn't collapse the company — Recursion still has one of the broadest pipelines in the sector — but it put a bracket around what "AI-discovered" actually guarantees. The answer, so far, is: a faster path to the clinic, not a safer one once you're there.

Schrödinger plays a different game. Its compound zasocitinib, developed through a Takeda partnership, has Phase III data in hand and represents the nearest-term shot at an approval. Schrödinger's edge is its physics-based simulation platform, which sits upstream of pure machine learning — closer to computational chemistry than to transformer models, though the distinction is blurring fast.

The 60-Percent Question

Analysts who track the pipeline closely put the probability of a first FDA approval from an AI-discovered compound at roughly 60 percent by end of 2027. That framing matters: it's not a certainty, and it's not a long shot. It's a coin flip with slightly better odds, stretched over 18 months.

The structural bottleneck isn't model quality. It's data scarcity at the clinical end. AI systems learn from what exists, and what exists in drug development is an enormous archive of failure — trials that didn't work, targets that didn't pan out, populations that didn't respond. Training on that data makes models good at avoiding known dead ends. It doesn't make them good at predicting novel biological mechanisms in humans, because humans in trials don't behave like molecules in simulation.

That gap — between digital prediction and biological reality — is where the next five years of AI biotech will be decided.

The ITIF published a brief this month arguing that policymakers need to catch up: FDA review frameworks, data exclusivity rules, and clinical trial design standards were written for a world where drug discovery was slow and linear. AI-native pipelines are fast and iterative. The regulatory surface hasn't adapted. Companies like Insilico are navigating agencies that don't have official guidance on what "AI-discovered" even means for review purposes.

From Pilots to Operating Models

Beyond the individual pipeline stories, something structural is shifting inside large pharma. The 2026 Benchling Biotech AI Report — surveying R&D teams across major biopharma — found that the industry has exited the pilot era. Half of organizations adopting AI already report faster time-to-target. Forty-two percent see measurable improvement in hit rates. The adoption language has changed from "we're experimenting with" to "we've rebuilt the data environment around."

That's not a triumph — it's a transition. The companies outperforming are the ones that treated AI integration as an infrastructure problem, not a software purchase. They rebuilt data pipelines, unified assay records, hired computational biologists who can talk to ML engineers. The ones underperforming are running the same fragmented data environments they had in 2022, with a model bolted on top.

Y Combinator's latest cohort includes multiple drug discovery startups targeting specific niches — rare diseases, neglected tropical diseases, metabolic disorders — where the small-molecule problem space is bounded enough for current models to be genuinely useful. That's where the near-term wins are most likely to come from: not the ambitious shots at cancer or neurodegeneration, but the constrained problems where AI has actual leverage.

---

Two hundred candidates. Zero approvals. One year, maybe two, before that second number changes or the narrative buckles under the weight of unmet expectations. Either outcome will be clarifying. The biology doesn't care about the pitch deck, and the trial data is coming regardless of what runs on the GPUs upstream. The most interesting thing about AI biotech in 2026 isn't the models — it's that we're finally close enough to find out if any of it works.