AI Agents in Scientific R&D: AlphaFold, Insilico, and the Reproducibility Bar

Drug discovery and protein science are uniquely hard cases for evaluating AI agents. The feedback loop is measured in years, not minutes: a molecule designed in 2024 reads out a phase II trial around 2027 at the earliest. That makes vendor claims cheap and verification expensive.

For this lesson we apply a strict bar: we only count evidence that survives at least one of the following filters — (1) peer review in Nature, Science, or a comparable journal; (2) regulatory filings (SEC 10-Ks, EU EMA dossiers, FDA INDs); (3) public clinical-trial registries (clinicaltrials.gov). Press releases, conference keynotes, and analyst reports do not count on their own.

The strongest evidence by far is in protein structure prediction. The weakest evidence is in end-to-end drug discovery. We'll work through both, plus what sits in between, and what the 2024 Nobel Prize in Chemistry does and does not signal.

The anchor citation is Jumper et al., "Highly accurate protein structure prediction with AlphaFold", Nature 596, 583–589 (2021), DOI 10.1038/s41586-021-03819-2. The companion paper for the database is Varadi et al., Nucleic Acids Research, 2022.

What the paper actually established, with measured benchmarks:

• Median backbone accuracy of 0.96 Å RMSD on CASP14 free-modelling targets, against ~2.8 Å for the next best method.
• For ~58% of CASP14 targets, AlphaFold's prediction was indistinguishable from experimental structures within their resolution.
• The CASP14 result (December 2020) is what convinced the structural biology community — CASP is a blind competition; targets are released as sequences and scored against unreleased experimental structures.

This is the cleanest possible evidence: a community-run blind benchmark, peer-reviewed result, reproducible code. The result also generalized — that's why it survived.

DeepMind and EMBL-EBI released the AlphaFold Protein Structure Database in 2021 and expanded it to roughly 200 million predicted structures in 2022 — effectively every catalogued protein from UniProt. It is hosted by EMBL-EBI (a public European institution), not by Google.

Usage signal that is hard to fake: AlphaFold structures are now routinely cited in structural biology papers as the starting point when no experimental structure exists. The follow-on AlphaFold 3 paper (Abramson et al., Nature, 2024) extended the model to protein-ligand, protein-DNA, and protein-RNA complexes — the building blocks for drug discovery.

This matters for the business question because the data moat here is real and free. Pharma R&D teams worldwide are using these structures whether they pay anyone or not. That is unusually unambiguous evidence of value, but note what it is not: it is not evidence that drugs designed against these structures are reaching patients faster.

On 9 October 2024, the Royal Swedish Academy of Sciences awarded the Nobel Prize in Chemistry to David Baker (one half, for computational protein design) and jointly to Demis Hassabis and John M. Jumper (one half, for AlphaFold). The official prize page is at nobelprize.org/prizes/chemistry/2024.

A Nobel is not a substitute for empirical evidence, but as a signal it carries weight: the Royal Swedish Academy is conservative, the chemistry committee is independent of industry, and the citation is unusually specific — it credits both the predictive breakthrough (AlphaFold) and the generative one (Baker's work on de novo protein design via RoseTTAFold and related tools).

What the prize does not establish: that AI agents are economically transformative across business. It establishes that a specific class of AI — supervised models trained on the Protein Data Bank — solved one of biology's oldest problems. Be careful not to generalize beyond that.

In November 2021 Alphabet announced Isomorphic Labs, a London-based subsidiary spun out of DeepMind and led by Demis Hassabis, with the stated mission of "reimagining the entire drug discovery process from first principles with an AI-first approach."

What is verifiable from Alphabet's public disclosures (10-Ks, earnings calls): Isomorphic has signed research collaboration deals with Eli Lilly (announced January 2024) and Novartis (announced January 2024), with upfront payments and milestones disclosed in the partners' filings (Lilly's deal involved an upfront payment of $45 million, with up to $1.7 billion in milestones; Novartis $37.5 million upfront, up to $1.2 billion). These are real cash flows on real balance sheets.

What is not yet verifiable: that any Isomorphic-originated molecule has entered the clinic. Drug pipelines take 10+ years. Treat the deals as evidence of belief by sophisticated buyers; treat them as not yet evidence of clinical or commercial success.

Hong Kong-based Insilico Medicine is the most-cited example of an AI agent platform producing a clinical candidate. Their lead molecule, rentosertib (INS018_055), an inhibitor of TNIK for idiopathic pulmonary fibrosis (IPF), entered phase II trials. The trial is registered on clinicaltrials.gov — search for the molecule code; specific NCT identifiers have been disclosed in company press releases and tracked by independent outlets including Nature Biotechnology and FierceBiotech.

Claims to be careful with:

• "AI-designed" usually means AI-assisted target identification and lead optimization, not autonomous discovery. Chemists are deeply involved at every step.
• "Time to IND" claims (Insilico has cited ~18 months versus an industry baseline of ~4 years) compare apples to oranges — molecule classes, target novelty, and trial design all vary.
• The candidate has not yet read out a pivotal trial. Phase II success rates industry-wide are around 30%; phase III about 60%.

The right framing: Insilico is the strongest public proof-of-concept that AI-augmented discovery can produce a clinical candidate. It is not yet proof that the resulting drugs work better, faster, or cheaper than the alternative.

Recursion Pharmaceuticals (NASDAQ: RXRX) operates a different model: high-content imaging of cellular phenotypes across millions of perturbations, with deep learning to find drug-like signals. Their 10-K filings (search SEC EDGAR for ticker RXRX) disclose pipeline size, partnership economics, and R&D burn.

What the filings actually show (per their FY2024 10-K, filed 2025):

• A pipeline of clinical and preclinical programs, mostly in rare/oncology indications.
• Strategic deals with Roche/Genentech and Bayer, with upfront payments and milestones.
• A 2024 acquisition of Exscientia (another AI-first biotech) — the deal closed in November 2024, consolidating two of the largest AI-discovery balance sheets.
• Negative net income, as is normal for clinical-stage biotech.

Recursion's value is being tested in public: it is a listed company whose share price reflects market belief about its AI platform. As of January 2026, that share price has been volatile and well below its 2021 SPAC-merger peak. Markets are still adjudicating.

The pipeline, and which steps have the strongest AI evidence:

Read the pipeline diagram left-to-right. The further right you go, the weaker the AI evidence:

• Target identification, structure prediction, virtual screening, lead optimization: strong evidence. AlphaFold is the proof point for structure; large benchmark datasets (e.g., PoseBusters, Nature Methods 2024, for docking realism) keep this honest.
• Preclinical: mixed evidence. Animal-model predictivity is the bottleneck regardless of how the molecule was found.
• Clinical: essentially no AI-specific evidence. Trials run the same way regardless of how the molecule was designed. The question "do AI-designed drugs work?" cannot be answered with current data — the molecules have not aged enough.
• Regulatory and patient outcomes: no evidence yet, by definition.

The honest summary in 2026: AI has demonstrably changed the upstream economics of drug discovery (cheaper structures, more candidates per dollar). Whether it changes downstream success rates is the open empirical question — and we will not know for several more years.

Practical filters when a vendor, founder, or analyst makes a claim about AI agents in R&D:

• Look for the registry. "In the clinic" should mean a real NCT identifier you can read. If it doesn't, it isn't.
• Look for the comparator. "50% faster" relative to what baseline? Was the baseline the same target class, same modality, same indication?
• Look for the journal. Peer-reviewed papers in Nature, Science, Cell, Nature Methods, Nature Biotechnology, JACS carry weight. Conference abstracts and preprints carry less.
• Look for the partner's filings. When a biotech says "we have a partnership with Big Pharma X," the upfront-and-milestone numbers are usually in Big Pharma X's 10-K or 20-F. Cross-reference.
• Look at the dates. The pipeline takes a decade. A 2022 announcement of an AI-designed candidate will not have phase III data before ~2030.

If a claim survives all five filters, it is probably real. If it dies at the first filter, treat it as marketing.