Uthando Nethemba

Recursion and Exscientia: the phenotypic engine can discover, but the clinic has not confirmed it

—

by

itembalabantu
in

Evidence-first notes on bioscience and deep tech, at the edge of the lab and the market. Information only — not investment advice.

The 30-second version

  • What. Recursion’s phenotypic engine (Recursion OS, a phenomics data universe of more than 25 petabytes) has demonstrated a real capability: to surface candidates from cellular phenotypes without knowing the target up front. But the assets that actually reached later-stage clinic — REC-4881 (MEK1/2) and REC-617 (CDK7) — are both previously-validated targets on known chemistry, while the pure phenotypic-origin assets (REC-994, REC-2282, REC-3964) were largely halted in May 2025. The gap between “discovery is demonstrated” and “clinical translation is not” shows up most sharply here.
  • So what. “Recursion OS validated the clinic” collapses into a misleading one-liner. There are zero approved drugs from this phenotypic route. The one positive signal, REC-4881 in familial adenomatous polyposis (FAP), is a single-arm study with a tiny sample (efficacy-evaluable 6 then 12), a surrogate endpoint (polyp burden), and an outlier included (one patient +595%). This is not a “positive Phase 2” headline; it is preliminary Phase 1b/2 data. Every quantitative claim below is attributed to company IR, press or preprint, not independently confirmed as a hard clinical outcome.
  • Now what. Treat “we can discover it” as separate from “it works in patients.” Exscientia — acquired by Recursion in November 2024 — is the origin of the “first AI-designed drug in the clinic” narrative, and it demonstrated discovery speed (DSP-1181 designed in ~12 months, EXS-21546 in ~8 months) but both flagship assets were discontinued (DSP-1181 after Phase 1; EXS-21546’s A2A target dropped in 2023). Notably, the merger’s first concrete move was pruning, not expansion: four programs halted. Verification is PARTIAL (14 confirmed / 0 refuted / 4 unverified); the sharpest attribution questions are finalized in later parts.

[demo-gap note] A discovery-layer capability is not a clinical outcome. Recursion OS’s 25 PB phenomics, Phenom-1 (3.5 billion+ cell images) and BioHive-2 (~4,800 NVIDIA GPUs, >1.2 exaFLOPS) demonstrate mapping of perturbation-to-phenotype at scale — a discovery/preclinical layer result, not evidence of clinical efficacy. REC-4881’s “positive” preliminary data is a single-arm, small-sample, surrogate-endpoint readout. The phenotypic engine’s unique contribution to clinical success cannot be attributed to REC-4881 or REC-617, because both are already-validated targets. “Can discover” is demonstrated; “translates to the clinic” is not.

The five-minute read

The one question: does phenotypic discovery translate into clinical efficacy?

Recursion and the AI-drug field get compressed into headlines about compressing decades of biology or designing drugs in months. One layer in, the useful question is whether the phenotypic engine’s discovery capability translates into hard clinical outcomes. The firm’s phenotypic-track finding is that it has not yet been shown: the discovery layer is demonstrated, but the assets that survived to later clinic are the ones with previously-validated targets, while the pure phenotypic-origin programs were the ones that were cut.

One positive signal versus four halts

Recursion OS — what it demonstrates. The phenomics data universe exceeds 25 petabytes of biological images and high-dimensional readouts; the Phenom-1 foundation model was trained on 3.5 billion+ cell images on BioHive-1; BioHive-2 runs on ~4,800 NVIDIA GPUs at >1.2 exaFLOPS (the company’s claim of the largest supercomputer owned by a pharma company). This methodologically demonstrates the ability to relate compounds and genetic perturbations by phenotypic similarity — a discovery-layer capability, distinct from clinical efficacy prediction.

The positive signal — REC-4881 (MEK1/2, FAP). In the single-arm TUPELO Phase 1b/2, preliminary data (2025-05-04, cutoff 2025-03-17) showed a median polyp-burden reduction of −43% at Week 13 in 6 efficacy-evaluable patients, with 5/6 (83%) reduced by −31% to −82% and one patient (17%) increased +595% (an outlier). A 2025-12-08 update (cutoff 2025-11-25) reported n=12, median −43% at 12 weeks; at Week 25 (12 weeks after stopping dosing), 82% (9/11) maintained a reduction, median −53%. Safety was mostly Grade 1–2 (Grade 3 15.8%, no Grade ≥4), consistent with the MEK-inhibitor class. The caveats are mandatory: tiny sample, no control arm, an included outlier, and a surrogate endpoint. MEK1/2 is not a new target, so the phenomics contribution is better read as indication mapping (FAP) than novel-target discovery — and REC-4881’s molecular origin (in-house design vs. in-license) is unverified.

The four halts — the May 2025 portfolio reset (the merger’s first move). REC-994 (superoxide scavenger, cerebral cavernous malformation), REC-2282 (pan-HDAC, NF2-mutant meningioma), REC-3964 (C. difficile TcdB toxin inhibitor, CDI) and REC-4209 (undisclosed, idiopathic pulmonary fibrosis, preclinical) were discontinued. The first three carry the character of phenotype-driven discovery assets; the assets that survived and produced a clinical signal — REC-4881 (MEK) and REC-617 (CDK7) — are previously-validated targets. That is the data behind the through-line.

[diagram: phenotypic discovery vs. clinical translation — the regression to validated targets]

  Track                    Asset(s)              Status              Target type
  ─────────────────────────────────────────────────────────────────────────────
  Survived to clinic       REC-4881 (MEK1/2)     signal, tiny n      VALIDATED
  (efficacy signal)        REC-617  (CDK7)        early activity      VALIDATED
  ─────────────────────────────────────────────────────────────────────────────
  Pure phenotypic-origin   REC-994  (CCM)         HALTED 2025-05      diverse /
  (target-agnostic start)  REC-2282 (NF2)         HALTED 2025-05      mechanism-
                           REC-3964 (CDI)         HALTED 2025-05      driven
  ─────────────────────────────────────────────────────────────────────────────
  Exscientia "AI-first"    DSP-1181 (5-HT1A)      HALTED after Ph1    speed shown,
  (design-speed narrative) EXS-21546 (A2A)        HALTED 2023         clinic failed
  ─────────────────────────────────────────────────────────────────────────────
  approved drugs from the phenotypic route:  0
  through-line:  discovery is demonstrated  !=  clinical translation is confirmed
The assets that reached later clinic converge on validated targets. Pure phenotypic-origin programs were the ones halted in May 2025; the surviving efficacy signals sit on known targets (MEK, CDK7). REC-4881’s signal is single-arm, small-sample, surrogate-endpoint. Exscientia’s flagship AI-designed assets demonstrated discovery speed but were discontinued. Figures are attributed to company IR/press, not independently confirmed.

Deep dive

1. Background — porting the “real bottleneck” lens to AI drug discovery

This Part 1 transplants the series’ “the real bottleneck next to the headline” lens onto phenotypic AI-drug platforms (Recursion plus the Exscientia it absorbed in 2024). The headline is always the scale of the phenomics engine or the speed of AI design; the firm’s question is whether that discovery capability translates into hard clinical outcomes. Verification status is PARTIAL — merger terms, pipeline readouts and halts were cross-checked across Recursion IR, GlobeNewswire, SEC 6-K and trade press, but the full adversarial pass on attribution boundaries is finalized in later parts.

2. What this part newly establishes

  • Phenomics demonstrates target-agnostic discovery [company/press, discovery layer]: 25 PB, Phenom-1, “maps of biology/chemistry” support surfacing candidates without a known target. This is a methodological demonstration at the discovery/preclinical layer only.
  • The clinical signal is one asset, and it is fragile [press/IR, cautious]: REC-4881 FAP is single-arm, n=6 then 12, median −43%, outlier included; a surrogate endpoint that would need large, randomized, long-term confirmation to support approval.
  • Pure phenotypic-origin assets underperformed [GEN/IR, confirmed]: REC-994, REC-2282 and REC-3964 were halted in May 2025, mostly for “absence of a clear signal.”
  • The Exscientia “AI-first in clinic” narrative shows speed, not clinical success [press, confirmed]: DSP-1181 (5-HT1A agonist, OCD; Sumitomo Dainippon Pharma collaboration) entered Phase 1 on 2020-01-30 with a “discovered in 12 months vs. the usual 5 years” narrative, then was discontinued after Phase 1 (company: expected standards not met; safety fine). EXS-21546 (A2A antagonist, immuno-oncology; Evotec collaboration), designed in ~8 months, was halted in 2023 when sufficient A2A receptor occupancy within the therapeutic window proved difficult. Both failed on target biology/pharmacology, not on design speed.
  • The merger’s first move was pruning [SEC/press, confirmed]: Recursion acquired all shares of Exscientia (announced 2024-08, value ~$650M–688M depending on the share price at announcement; 0.7729 RXRX per Exscientia share; post-merger ownership ~74% Recursion / ~26% Exscientia; closing ~2024-11-20). Chris Gibson stayed CEO; David Hallett became Recursion CSO. The first concrete integration action was the four halts above.

3. Strengths and limits of the methodology

Strengths (discovery layer): target-agnostic discovery — retracing candidates, indications and mechanism hypotheses from phenotypic similarity alone, favorable for exploring novel biological space (the FAP indication mapping for REC-4881 is an example); and economies of scale — 25 PB, 3.5 billion images and 1.2 exaFLOPS to map perturbation-phenotype relationships in an automated wet-lab loop. Limits (clinical-translation layer): the cost of being mechanism-agnostic — when target/mechanism is uncertain, dose-response design, biomarkers and patient selection are weaker, which fits the pattern of the four halts ending for “absence of a clear signal”; the phenotype-to-patient gap — a cellular-image phenotype is not guaranteed to map onto a human disease endpoint (in vitro-to-clinic gap); and the regression phenomenon — the assets that survived to later clinic are validated targets (MEK, CDK7), so the phenotypic engine’s distinctive strength (target-agnostic discovery) cannot be credited as the direct cause of clinical success.

4. Neighbouring domains

The cross-domain hook is methodological rather than forced. This is the clinical version of the demo-gap the computing and bio-FM series recorded: a large imaging/design platform (demonstrated at the design and preclinical layer) versus hard clinical outcomes. The capital-versus-evidence mismatch recurs too — 25 PB of phenomics and a 1.2-exaFLOPS supercomputer against zero approvals — echoing the bio-FM series. Whether these methods translate into CKM (cardio-renal-metabolic) or oncology hard outcomes is deferred; the through-line recorded now is the structural similarity, not a claimed clinical payoff.

5. Commercialization and market context (actors, platforms)

TRL sits at roughly 4–6 (multiple Phase 1/2 assets, zero approvals; REC-4881 in FAP is the front-line efficacy signal). This is the platform/asset angle; share-price implications are out of scope and escalated.

ActorTypeAssets / platformsCommercial signal (attributed)
RecursionListed (RXRX)Recursion OS (phenomics), Phenom-1, BioHive-2; REC-4881, REC-617, REC-1245, REC-3565, REC-7735, REV10225 PB phenomics; ~4,800-GPU BioHive-2 (company claim). Six active programs post-reset plus REC-4539 paused. Factual description only.
Exscientia (acquired)Formerly listed (EXAI)Generative / active-learning chemistry design; REC-617 (ex-GTAEXS617, CDK7)Acquired by Recursion, closing ~2024-11-20; 0.7729 RXRX per share; ~26% of the combined company. DSP-1181 and EXS-21546 both discontinued.
NVIDIAListedGPU / BioHive-2 infrastructurePowers BioHive-2 (~4,800 GPUs). Factual description only.
EvotecListedEXS-21546 collaboration (A2A)Co-development partner; program halted 2023. Factual description only.
Sumitomo PharmaListedDSP-1181 collaboration (5-HT1A)Co-development partner; program halted after Phase 1. Factual description only.

The company describes the merger logic as biology (phenomics) + chemistry (generative/structure-based design) = end-to-end, with expected synergies of ~$100M per year (company forward-looking, unverified precision). REC-617 (ex-Exscientia, CDK7) is the surviving “chemistry” asset in the pipeline: in ELUCIDATE Phase 1/2, one confirmed partial response in platinum-resistant ovarian cancer (6 months+) and 4 patients with stable disease across 18 response-evaluable, mostly Grade 1–2, zero treatment discontinuations (2024-12-09 interim) — real, but CDK7 is again a validated target. No single-winner conclusion is asserted; vendor commercial figures are attributed.

6. The skeptic’s bottom line

  • Verified-clean: merger terms and governance (2024-11; 0.7729; 74/26), REC-4881 TUPELO figures, the four halts (REC-994/2282/3964/4209), REC-617 interim, DSP-1181/EXS-21546 discontinuations, and Recursion OS scale — cross-checked against primary IR/SEC/press.
  • Proceed-with-caveats: the REC-4881 “positive” description carries mandatory conditions — single-arm, very small sample, surrogate endpoint, outlier included — and must not be expanded into confirmed efficacy. “Phenomics demonstrates discovery” is limited to the discovery layer and must not be extended to clinical efficacy. Statements about the listed company (RXRX) exclude any buy/sell implication and are kept separate from the investment layer.
  • Hold: (1) the complete narrative that “Recursion OS validated the clinic” — the flagship assets are validated targets and approvals are zero; (2) a “positive Phase 2” headline for REC-4881 — this is single-arm preliminary Phase 1b/2 data; (3) attributing phenotypic discovery’s clinical contribution to REC-4881 or REC-617 — inseparable because the targets are pre-validated.
  • Escalated: to skeptic (sample and attribution boundaries) and to Principal (neutral frame for the listed company).

7. What to watch (falsifiable predictions)

  1. REC-4881 (FAP) will not pass a hard or approvable endpoint in a randomized, controlled confirmatory trial within the next 24 months (the single-arm signal does not hold up against a control arm). Falsifiable by Recursion’s confirmatory-trial top-line disclosure.
  2. Recursion will not report a Phase 2 efficacy proof-of-concept in public data for a pure phenotypic-origin (target-agnostic start) asset; clinical wins will keep concentrating on validated targets (MEK, CDK7, MALT1). Falsifiable by a positive Phase 2 readout from a target-agnostic-origin asset.
  3. Despite the ~$100M synergy and pipeline-expansion narrative, at least one more program halt or priority change will occur within the next 18 months (the pruning-first pattern persists). Falsifiable by disclosure.

References

Source knowledge asset: knowledge-base/deep-dives/ai-drug-clinical-readout/part1-phenotypic-platforms.md (generated 2026-07-12, verification PARTIAL — 14 confirmed / 0 refuted / 4 unverified). This draft inherits the figures and source attributions of the original part and creates no new figures, numbers or sources.

Disclosure

This post is for information only and is not investment advice. The author holds no position in, and no financial interest in, the companies mentioned (Recursion RXRX, Exscientia, NVIDIA, Evotec, Sumitomo Pharma, and others).

COI note: This document describes a listed company (Recursion, RXRX), an acquired formerly-listed company (Exscientia), and collaborators (NVIDIA, Evotec, Sumitomo Pharma) in a factual, neutral technology/platform context. It includes negative facts — pipeline halts, discontinuations, safety events — but implies no ranking or buy/sell judgment. Company materials (IR, synergy estimates) are attributed and cross-checked against clinical registrations and SEC filings. Quantitative claims are attributed as “vendor/company/press/preprint claims” and are not independently verified as hard clinical outcomes; specifically, the merger value ($650M vs. $688M), REC-4881’s molecular origin and the precise boundary of the phenomics contribution, the ~$100M annual synergy, and the current cash runway remain unverified.

Comments

Leave a comment