Explore the research

Drug discovery

Can AI help create treatments that prove safe and useful in people, with less wasted effort?

20%

field summary estimate

6 outcomes · Assessed November 7, 2026 · Low confidence

The field estimate averages its outcome scores. It is a reading aid, not a measured fraction of the field that AI manages.

How the estimate built up.

Each band is one outcome's share of the field estimate, so the top line is the estimate itself. Every field starts at 0% on December 31, 2022. Dates before the latest are retrospective estimates. Point at a band to see that outcome's score on each date.

Translation20%
Target validation25%
Portfolio productivity10%
Development efficiency25%
Clinical benefit15%
Candidate quality25%

View exact historical values and reconstruction notes

What we measure

Each percentage is a judgment about progress toward independent advancement. Open the evidence to inspect the basis for it.

Candidate quality

25%estimated progress

What we track

Fraction of candidates meeting prespecified potency, selectivity, exposure, and toxicity requirements in relevant assays.

Why this estimate?

One experimentally supported candidate demonstrates useful design assistance, but complete proposal-to-qualified-candidate rates and self-managed validation are unknown.

Advancements

Absci zero-shot designed antibodies include three tighter binders than trastuzumabPosted March 29, 2023 (version 3). Absci screened more than a million AI-generated anti-HER2 antibody variants, reporting binding rates of 10.6% for HCDR3 designs and 1.8% for HCDR123 designs, about four and eleven times the rates for sequences randomly sampled from the Observed Antibody Space. Of 421 binders characterized by surface plasmon resonance, three bound more tightly than the therapeutic antibody trastuzumab.
AI-guided TNIK inhibitor shows antifibrotic activity2024 publication. AI-guided TNIK inhibitor INS018_055 showed drug-like properties and antifibrotic activity in preclinical models.
Generative model yields 6 active compounds from 22 synthesized against MRSAAugust 15, 2025. In work published in Cell, generative algorithms designed more than 29 million compounds against Staphylococcus aureus, narrowed to about 90; of 22 synthesized and tested, six showed strong activity against multi-drug-resistant S. aureus in vitro, and the lead, DN1, cleared an MRSA skin infection in a mouse model. A parallel fragment-based run produced NG1, effective against drug-resistant gonorrhea in mice.

Read the dated findings and sources

Status: Validated candidate example; cohort pass fraction unknown. Sources accessed November 7, 2026. This is a documentary baseline, not an independent replication or an assertion of the worldwide best result. An unknown measurement is not zero.

Dated findings

2024 publication: AI-guided TNIK inhibitor INS018_055 showed drug-like properties and antifibrotic activity in preclinical models. TNIK inhibitor discovery, Nature Biotechnology.

Interpretation and tracking value

One selected candidate does not reveal the fraction of all proposals meeting potency, selectivity, exposure and toxicity thresholds. Sponsor participation and selection effects matter.

Next comparable measurement

Record all nominated candidates and prespecified assay thresholds; report each requirement and joint pass fraction, including failures. Preserve this dated observation; append future results with their model/system version, denominator, budget and evidence level rather than overwriting it.

Open the full research note

Clinical benefit

15%estimated progress

What we track

Prespecified patient-relevant primary outcomes, effect sizes, uncertainty, adverse events, and withdrawals in controlled trials.

Why this estimate?

A small randomized trial provides an early clinical signal. Exploratory efficacy, human-led trials and missing confirmatory benefit keep autonomy progress low.

Advancements

Recursion's REC-994 meets phase 2 safety endpoint but shows no symptom improvementSeptember 3, 2024. In the 62-participant SYCAMORE phase 2 trial in symptomatic cerebral cavernous malformation, REC-994 met its primary safety and tolerability endpoint with similar adverse events across placebo and the 200 mg and 400 mg arms at 12 months. Exploratory MRI measures showed only a trend toward reduced lesion volume and hemosiderin ring size at 400 mg, and patient- or physician-reported outcomes showed no improvement at 12 months; the study was not powered for statistical significance.
Rentosertib improves lung function in a phase 2a trialJune 3, 2025. In 71 patients over 12 weeks, 60 mg daily rentosertib showed mean FVC change +98.4 mL (95% CI 10.9–185.9), versus -20.3 mL (-116.1–75.6) with placebo.
Rentosertib enters phase IIIJuly 7, 2026. The sponsor announced phase III initiation in July 2026; initiation is not a phase III efficacy result.

Read the dated findings and sources

Status: Small randomized trial; efficacy exploratory. Sources accessed November 7, 2026. This is a documentary baseline, not an independent replication or an assertion of the worldwide best result. An unknown measurement is not zero.

Dated findings

June 3, 2025: In 71 patients over 12 weeks, 60 mg daily rentosertib showed mean FVC change +98.4 mL (95% CI 10.9–185.9), versus -20.3 mL (-116.1–75.6) with placebo. Rentosertib randomized phase 2a trial.

July 7, 2026: The sponsor announced phase III initiation in July 2026; initiation is not a phase III efficacy result. Insilico: phase III initiation announcement.

Interpretation and tracking value

Safety was the primary endpoint; these are arm-specific intervals, not the interval for the treatment difference. The study does not establish survival benefit or confirmatory efficacy.

Next comparable measurement

Track a prespecified confirmatory clinical endpoint, treatment-effect interval, adverse events, withdrawals and longer follow-up; retain this as phase 2a evidence. Preserve this dated observation; append future results with their model/system version, denominator, budget and evidence level rather than overwriting it.

Open the full research note

Development efficiency

25%estimated progress

What we track

Elapsed months, total dollars, and human effort from fixed program start to candidate nomination and each trial phase; include failed programs.

Why this estimate?

A reported nomination timeline supports faster bounded discovery; failure-adjusted portfolio costs and autonomous clinical development remain unestablished.

Advancements

TNIK drug candidate nominated in about 18 months2024 publication. The authors report roughly 18 months from target discovery to preclinical candidate nomination.
Boltz-2 computes binding affinities over 1000 times faster than FEP simulationsbioRxiv posting dated 2025-06-18. MIT's Jameel Clinic reports that Boltz-2 runs more than 1000 times faster than physics-based free-energy perturbation simulations while approaching their accuracy, and releases the code, weights and training pipeline under an MIT license permitting commercial use.
Insilico reports 12 to 18 months from program start to preclinical candidateDecember 10, 2025. Announcing ISM3830, described as its 23rd AI-empowered preclinical candidate since 2021, Insilico Medicine reported an average of 12 to 18 months per program from project initiation to preclinical candidate nomination, with 60 to 200 molecules synthesized and tested per program, against a stated 2.5 to 4 years for traditional early-stage discovery.

Read the dated findings and sources

Status: Measured single-program timeline. Sources accessed November 7, 2026. This is a documentary baseline, not an independent replication or an assertion of the worldwide best result. An unknown measurement is not zero.

Dated findings

2024 publication: The authors report roughly 18 months from target discovery to preclinical candidate nomination. TNIK inhibitor discovery, Nature Biotechnology.

Interpretation and tracking value

This excludes later clinical development and is not a failure-adjusted portfolio cost or causal comparison with conventional discovery.

Next comparable measurement

Track fixed program-start dates, staffing and full expenditure through each phase; include discontinued programs and matched indications/modalities. Preserve this dated observation; append future results with their model/system version, denominator, budget and evidence level rather than overwriting it.

Open the full research note

Portfolio productivity

10%estimated progress

What we track

Independently demonstrated beneficial treatments per total portfolio dollar and year, accounting for attrition and delayed outcomes.

Why this estimate?

An observable clinical pipeline is progress toward translation, but beneficial treatments per full portfolio dollar and year are not yet established by the baseline.

Advancements

Recursion ends three clinical programmes and pauses an oncology candidateMay 6, 2025. Six months after merging with Exscientia, Recursion ended development of REC-2282 in neurofibromatosis type 2 and REC-994 in cerebral cavernous malformation, sought out-licensing for REC-3964 in C. difficile infection, paused the clinical-stage oncology candidate REC-4539 and cut an undisclosed preclinical programme, leaving six active programmes, four in oncology and two in rare disease, none in late-stage trials.
Cohort study follows AI-enabled drugs through clinical phases2026 publication; cohort cutoff March 1, 2026. The clinical pipeline analysis reports substantially different success rates between early phases.

Read the dated findings and sources

Status: Pipeline measurement; beneficial-treatment productivity unknown. Sources accessed November 7, 2026. This is a documentary baseline, not an independent replication or an assertion of the worldwide best result. An unknown measurement is not zero.

Dated findings

2026 publication; cohort cutoff March 1, 2026: The clinical pipeline analysis reports substantially different success rates between early phases. AI-enabled clinical pipeline cohort study.

Interpretation and tracking value

Phase passage is not an independently demonstrated beneficial treatment, and the study does not establish treatments per total portfolio dollar and year.

Next comparable measurement

Follow all portfolio entrants through confirmatory benefit and discontinuation; count all spending and elapsed time, including unsuccessful programs. Preserve this dated observation; append future results with their model/system version, denominator, budget and evidence level rather than overwriting it.

Open the full research note

Target validation

25%estimated progress

What we track

Proposed targets with independently reproduced disease-relevant experimental support / all targets selected for validation.

Why this estimate?

Disease-relevant target evidence supports useful selection, while outside replication rates and autonomous experimental validation remain incomplete.

Advancements

Deep learning screen of 7,500 molecules yields abaucin acting on LolE2023. Reported in Nature Chemical Biology, a neural network trained on growth-inhibition data from about 7,500 molecules screened against Acinetobacter baumannii identified abaucin, a narrow-spectrum antibacterial that disrupts lipoprotein trafficking through a mechanism involving LolE and controlled an A. baumannii infection in a mouse wound model.
AI-selected target TNIK is backed by experimental evidence2024 publication. The TNIK study connects AI target selection to experimental antifibrotic evidence.
AI-designed compound NG1 implicates LptA as a new antibacterial targetAugust 15, 2025. Follow-up work on NG1, the lead compound from a generative design run against Neisseria gonorrhoeae, found that it interacts with LptA, described as a novel drug target involved in synthesis of the bacterial outer membrane. NG1 killed N. gonorrhoeae in vitro and in a mouse model of drug-resistant infection.

Read the dated findings and sources

Status: Disease-relevant target evidence; independent hit rate unknown. Sources accessed November 7, 2026. This is a documentary baseline, not an independent replication or an assertion of the worldwide best result. An unknown measurement is not zero.

Dated findings

2024 publication: The TNIK study connects AI target selection to experimental antifibrotic evidence. TNIK inhibitor discovery, Nature Biotechnology.

Interpretation and tracking value

This is evidence for one target, not independently reproduced targets/all selected targets. Clinical activity also cannot by itself isolate the target-discovery method's contribution.

Next comparable measurement

Publish the full target queue and experimental failures; require outside replication in disease-relevant systems before counting validated targets. Preserve this dated observation; append future results with their model/system version, denominator, budget and evidence level rather than overwriting it.

Open the full research note

Translation

20%estimated progress

What we track

Phase-transition success by indication and modality, using complete cohorts, explicit follow-up, and time-to-event methods for ongoing programs.

Why this estimate?

Clinical phase-transition data provide measurable downstream progress. Heterogeneous cohorts and human-managed development prevent interpreting those rates as autonomous drug development.

Advancements

First pipeline analysis puts AI-discovered molecules at 80-90% phase 1 successFirst published online April 30, 2024. An analysis of AI-native biotech clinical pipelines in Drug Discovery Today reports an approximately 80-90% phase 1 success rate for AI-discovered molecules, well above historical industry averages, but an approximately 40% phase 2 success rate, in line with historical averages and based on a small sample.
AI-enabled drugs pass phase 1 at 77.8% and phase 2 at 19.5%2026 publication; cohort cutoff March 1, 2026. The study's cohort, with cutoff March 1, 2026, reports phase 1 and phase 2 success rates of 77.8% and 19.5%, respectively.

Read the dated findings and sources

Status: Measured descriptive clinical cohort. Sources accessed November 7, 2026. This is a documentary baseline, not an independent replication or an assertion of the worldwide best result. An unknown measurement is not zero.

Dated findings

2026 publication; cohort cutoff March 1, 2026: The study's cohort, with cutoff March 1, 2026, reports phase 1 and phase 2 success rates of 77.8% and 19.5%, respectively. AI-enabled clinical pipeline cohort study.

Interpretation and tracking value

Cross-sectional, heterogeneous AI-enabled assets are not a randomized comparison with conventional discovery. Definitions, pending outcomes and censoring prevent direct comparison with differently selected earlier cohorts.

Next comparable measurement

Track inception cohorts by indication and modality; archive phase transitions and failures, define success consistently and use time-to-event methods. Preserve this dated observation; append future results with their model/system version, denominator, budget and evidence level rather than overwriting it.

Open the full research note