Article
Evidence Infrastructure: What Two Weeks of AI Biotech News Actually Tell Us
Six AI-in-biotech stories landed in the last two weeks. Read in isolation they look like unrelated headlines. Read as a sequence, one pattern jumps out.
The industry has stopped asking whether AI works in biotech. It has started asking whether AI claims can be proven.
Discovery platforms are being measured on revenue cadence, not narrative. Regulators are quietly building the data plumbing for AI-mediated approvals. Diagnostics is shipping cleared products. Foundation-model claims are getting pressure-tested in peer review. Late-stage capital is underwriting AI-designed biologics against payer endpoints.
The throughline is evidence infrastructure. Here is what happened and why each piece matters.
1. FDA launches Real-Time Clinical Trial pilot
On April 28, FDA Commissioner Marty Makary and Chief AI Officer Jeremy Walsh announced two proof-of-concept Real-Time Clinical Trials. AstraZeneca's Phase 2 TRAVERSE study in mantle cell lymphoma and Amgen's Phase 1b STREAM-SCLC trial of tarlatamab will stream endpoints and safety signals directly to FDA via Paradigm Health's cloud platform. Walsh estimates 20 to 40 percent reductions in overall trial time are achievable. An RFI for a broader summer pilot is open until May 29.
This is the most consequential regulatory move of the year so far. Continuous submission has been a buzzword for a decade. It just became operational. The implication for clinical operations is direct: audit trails, lineage tracking, and AI-monitored safety signals are no longer aspirational architecture. They are a regulatory deliverable. Sponsors who treat their trial data infrastructure as a compliance afterthought will not be running these pilots.
2. Recursion misses Q1 revenue by ~60 percent
Recursion (RXRX) reported Q1 2026 results on May 6: $6.5M revenue against a $16.1M consensus, a $117.5M net loss, and $665M cash with runway into early 2028. Six days earlier, co-founder Chris Gibson announced he will not seek re-election to the board. Najat Khan is now running earnings calls as CEO.
The bellwether public TechBio is being valued on milestone revenue cadence, not platform aesthetics. This matters beyond one ticker. The "AI-native biotech" peer set, including Schrödinger, AbCellera, and Absci, will face the same gravity. Platform stories built on data scale and model novelty are no longer sufficient. The market wants evidence of pipeline progression and partnership economics that can be audited quarter by quarter.
3. Blackstone commits $250M to AI-engineered enzymes
On May 7, Blackstone Life Sciences became sole investor in a $250M structured deal funding Anagram Therapeutics through Phase 2 and commercial launch. The asset, ANG003, is a recombinant oral enzyme replacement for cystic fibrosis-related exocrine pancreatic insufficiency, designed via ATUM's ProteinGPS machine-learning platform.
Two things to notice. First, this is not a venture round. It is structured scale capital, the kind of vehicle PE uses for assets it believes can be underwritten against specific clinical and reimbursement milestones. Second, the AI-designed nature of the molecule did not disqualify the deal. It enabled it. Expect this template to repeat as more AI-engineered biologics reach IND and beyond.
4. Isomorphic Labs is "very close" to dosing humans
On April 25, Isomorphic Labs President Colin Murdoch confirmed the company is staffing up for first-in-human oncology trials of AI-designed candidates. IND clearance for ISM8969 landed January 28. The DeepMind spinout is reportedly in advanced talks to raise $2B-plus led by Thrive Capital, on top of last year's $600M round.
The next twelve months will deliver the first real test of whether generative drug design improves Phase 1 translation, or merely compresses preclinical timelines. That distinction is not academic. It determines whether AI in drug discovery is a productivity layer or a paradigm shift. Anyone making AI ROI claims internally should know which question their data actually answers.
5. Artera scores first FDA-cleared digital pathology AI for breast cancer
On May 6, Artera received 510(k) clearance for ArteraAI Breast, a multimodal AI risk-stratification test for early-stage HR+/HER2-negative invasive breast cancer. It joins recent clearances for PathAI's AISight Dx primary diagnosis platform and Ibex Prostate Detect.
Diagnostic AI is the quiet success story of 2026. While drug-discovery AI remains pre-revenue and platform-narrative dependent, diagnostic AI is shipping cleared products into actual reimbursement pathways. The reason is structural: diagnostic AI builds on a regulatory framework (SaMD) and an evidence pattern (sensitivity, specificity, head-to-head against pathologist consensus) that maps cleanly to existing FDA practice. Discovery AI lacks an equivalent evidence template, which is precisely the problem regulators and capital are now solving for.
6. Foundation-model reality checks hit bioRxiv
In early May, multiple preprints questioned single-cell foundation model performance. A zero-shot benchmarking study showed scFM embeddings underperform a highly-variable-gene baseline for trajectory inference, citing temporal compression artifacts. A separate paper introducing xVERSE argues transcriptomics-native architectures beat language-model ports by 17.9 percent.
The "more parameters, more biology" thesis is being empirically pressure-tested. For technical leaders evaluating build versus license decisions, the evidence is shifting toward domain-specific architectures and rigorous task-level benchmarking. Generalist bio-foundation models still have a role. They are not the default answer.
The regulatory backbone in the background
Underneath all of this sits the January 2026 FDA-EMA Joint Guiding Principles for Good Machine Learning Practice. The principles operationalize what auditable AI in drug development means: data lineage, model documentation, change control, performance monitoring, and human oversight as a continuous obligation, not a single submission artifact.
The Real-Time Clinical Trial pilot is the principles applied. The Recursion revenue conversation is the market asking whether the principles will be met. The PE investment in Anagram is capital betting that they can be. Read in this light, the two-week window stops looking like a news cycle and starts looking like a regulatory regime stabilizing in real time.
Why this matters for the next 18 months
The companies that win the next 18 months will not be the ones with the biggest models or the deepest pipelines on paper. They will be the ones whose AI claims survive auditable, regulator-grade documentation.
This is what we mean when we talk about evidence infrastructure. Not a buzzword. A specific operational capability: the ability to produce, on demand, the chain of evidence behind any AI-driven decision in a regulated workflow. Without it, even strong AI capabilities cannot graduate from pilot to production. With it, ordinary AI capabilities can become regulator-defensible products.
The companies in this report sit on different sides of that line. The ones moving toward it are the ones to watch.