AI drug discovery: revision 1 to 2

Revision 1

Revision 2

categories: ["foundations", "longevity"]

tags: ["machine learning", "drug development", "generative chemistry", "target discovery", "clinical trials", "pharmacology"]

tags: ["machine learning", "drug development", "generative chemistry", "target discovery", "clinical trials", "pharmacology"]

summary: "The use of machine learning to select drug targets, generate candidate molecules and predict failure, so far demonstrably faster at making compounds than at making medicines."

summary: "The use of machine learning to select drug targets, generate candidate molecules and predict failure, so far demonstrably faster at making compounds than at making medicines."

updated: "2026-07-27"

updated: "2026-08-22"

humanEvidence: "Several AI-derived candidates have reached phase 1 and phase 2 trials in people; as of mid-2026 none has completed phase 3 or been approved, and the largest clinical success is a repurposed existing drug."

humanEvidence: "AI-derived small molecules have reached phase 1 and 2 trials, with a first phase 3 begun in 2026; sponsors of an mRNA therapy built on algorithm-selected neoantigens report a phase 3 in melanoma met its endpoints, and no drug from either route is approved."

access: "Platforms are proprietary to pharmaceutical companies or licensed to them, and some models and datasets are public; no molecule designed this way is available outside a trial, though AI-suggested uses of approved drugs are."

access: "Platforms are proprietary to pharmaceutical companies or licensed to them, and some models and datasets are public; no molecule designed this way is available outside a trial, though AI-suggested uses of approved drugs are."

reversibility: "context"

issues: ["Needs a citable source for the absence of any approved AI-discovered drug.", "The phase 1 to approval rate is uncited; Cook 2014 supports the efficacy-failure point only."]

issues: ["Needs a citable source for the absence of any approved AI-discovered drug.", "The phase 1 to approval rate is uncited; Cook 2014 supports the efficacy-failure point only.", "The INTerpath-001 paragraph rests on a topline company announcement; detailed data are not yet published."]

---

```infobox

{ "label": "Type", "value": "Discovery and design method" },

{ "label": "Main tasks", "value": "Target choice, molecule generation, property prediction" },

{ "label": "First candidate in trials", "value": "2020" },

{ "label": "Furthest clinical stage", "value": "Phase 2" },

{ "label": "Furthest clinical stage", "value": "Phase 3 (first begun 2026)" },

{ "label": "Drugs approved from the approach", "value": "None as of 2026" },

{ "label": "Drugs approved from the approach", "value": "None as of August 2026" },

{ "label": "Main criticism", "value": "Attrition is a biology problem" },

{ "label": "Readiness", "value": "TRL 5" }

]

chemistry claims are the better supported ones: software now proposes synthesizable, potent-looking

compounds against a defined target at a speed no medicinal-chemistry team matches by hand, though how

much of that speed the models themselves explain is disputed. The clinical claims are not settled. As

of mid-2026 several AI-derived candidates have entered phase 1 and phase 2 trials, and none has

of August 2026 several AI-derived candidates have entered phase 1 and phase 2 trials, the first

completed phase 3 or been approved.

small molecule among them has begun phase 3, and none has been approved; the one positive phase 3

readout claimed for the field comes from a therapy in which the algorithmic step is antigen

selection rather than molecule design.

## What the models actually do

  { "year": "2020", "title": "Antibiotic found by screening model", "text": "The Collins laboratory identifies halicin using a graph neural network, the first widely cited antibiotic hit attributed to deep learning." },

  { "year": "2020", "title": "Antibiotic found by screening model", "text": "The Collins laboratory identifies halicin using a graph neural network, the first widely cited antibiotic hit attributed to deep learning." },

  { "year": "2021", "title": "AlphaFold2 released", "text": "Predicted structures for most of the human proteome are made public, changing what structure-based design can be attempted." },

  { "year": "2021", "title": "AlphaFold2 released", "text": "Predicted structures for most of the human proteome are made public, changing what structure-based design can be attempted." },

  { "year": "2021–2022", "title": "AI-derived target and molecule reach humans", "text": "Insilico's ISM001-055, a TNIK inhibitor for idiopathic pulmonary fibrosis, enters first-in-human testing after both target and compound were nominated computationally." },

  { "year": "2021–2022", "title": "AI-derived target and molecule reach humans", "text": "Insilico's ISM001-055, a TNIK inhibitor for idiopathic pulmonary fibrosis, enters first-in-human testing after both target and compound were nominated computationally." },

  { "year": "2024", "title": "Consolidation", "text": "Exscientia and Recursion merge, combining two of the largest AI-first drug discovery companies after neither had produced an approved medicine." }

  { "year": "2024", "title": "Consolidation", "text": "Exscientia and Recursion merge, combining two of the largest AI-first drug discovery companies after neither had produced an approved medicine." },

  { "year": "2026", "title": "First AI-designed small molecule enters phase 3", "text": "Insilico Medicine begins a 320-patient randomized placebo-controlled phase 3 trial of rentosertib in idiopathic pulmonary fibrosis in China, the furthest any generatively designed small molecule has gone." },

  { "year": "2026", "title": "Neoantigen therapy meets phase 3 endpoints", "text": "Merck and Moderna report that intismeran autogene, an individualized mRNA therapy encoding up to 34 algorithm-selected neoantigens, met its recurrence-free survival endpoint with pembrolizumab in resected melanoma." }

]

```

own platforms, which makes it the cleanest available test of the full pipeline. It cleared phase 1,

and results from a twelve-week randomized placebo-controlled phase 2a trial in China were published

in 2025: one dose arm showed a mean improvement in forced vital capacity where the placebo arm

declined.[^xu2025] That trial was not sized to establish clinical benefit, and the compound had not

declined.[^xu2025] That trial was not sized to establish clinical benefit. In July 2026 the company

entered phase 3 as of mid-2026.

announced the start of a 320-patient, 52-week randomized placebo-controlled phase 3 trial in China,

the first phase 3 for a small molecule whose target and structure were both machine-nominated; no

data from it exist.[^insilico2026]

The rest of the first wave has behaved like ordinary pharmacology. DSP-1181 was announced in 2020 as

the first AI-designed molecule to enter a clinical trial and did not progress beyond phase 1; a later

either. A designed nanoparticle scaffold is a component of a COVID-19 vaccine approved in South

Korea, but as of 2026 no de novo designed protein has been approved anywhere as a therapeutic drug.

A phase 3 result of a different kind arrived in August 2026, from outside small-molecule chemistry.

Merck and Moderna announced that intismeran autogene, an individualized neoantigen mRNA therapy

covered under [[personalized-mrna-cancer-vaccines]], met its primary endpoint of recurrence-free

survival and a key secondary endpoint of distant metastasis-free survival when added to

pembrolizumab in a 1,137-patient trial in resected stage IIB–IV melanoma. The claim is topline, from

an interim analysis, and detailed data had not been released as of late August 2026.[^merck2026]

Each course is built by sequencing the patient's tumour and encoding up to 34 of its mutations,

selected by what Moderna describes as a proprietary algorithm developed with Merck.[^moderna2023]

That computational step ranks which of a patient's existing mutations to encode rather than

generating a new chemical entity, so whether the result counts as an AI-derived drug succeeding in

phase 3 depends entirely on where the label's boundary is drawn.

> [!caution] What "AI-discovered" means

> No regulator recognizes the category, and companies apply the label to everything from a molecule

> drawn by a generative model to a conventional campaign in which a classifier filtered one plate.

[^zhavoronkov2019]: `paper` Zhavoronkov, A. et al. "Deep learning enables rapid identification of potent DDR1 kinase inhibitors." *Nature Biotechnology*, 2019. {DDR1 was a well-precedented target and the compounds resembled known inhibitors; chemists disputed how much of the speed the model explained.}

[^zhavoronkov2019]: `paper` Zhavoronkov, A. et al. "Deep learning enables rapid identification of potent DDR1 kinase inhibitors." *Nature Biotechnology*, 2019. {DDR1 was a well-precedented target and the compounds resembled known inhibitors; chemists disputed how much of the speed the model explained.}

[^stokes2020]: `paper` Stokes, J.M. et al. "A Deep Learning Approach to Antibiotic Discovery." *Cell*, 2020. {Halicin's activity was confirmed in mouse infection models; it has not entered human trials.}

[^stokes2020]: `paper` Stokes, J.M. et al. "A Deep Learning Approach to Antibiotic Discovery." *Cell*, 2020. {Halicin's activity was confirmed in mouse infection models; it has not entered human trials.}

[^xu2025]: `paper` Xu, Z. et al. "A generative AI-discovered TNIK inhibitor for idiopathic pulmonary fibrosis: a randomized phase 2a trial." *Nature Medicine*, 2025. {Seventy-one patients across three dose arms and placebo, treated for twelve weeks; a trial that size and length cannot establish clinical benefit in fibrosis.}

[^xu2025]: `paper` Xu, Z. et al. "A generative AI-discovered TNIK inhibitor for idiopathic pulmonary fibrosis: a randomized phase 2a trial." *Nature Medicine*, 2025. {Seventy-one patients across three dose arms and placebo, treated for twelve weeks; a trial that size and length cannot establish clinical benefit in fibrosis.}

[^insilico2026]: `statement` Insilico Medicine. "Insilico Initiates Phase III Clinical Trial for Rentosertib, Its AI-Empowered TNIK Inhibitor for Idiopathic Pulmonary Fibrosis." Press release, July 2026. {Announces trial initiation only, not results; a 320-patient, 52-week randomized placebo-controlled study at Chinese centres.}

[^merck2026]: `statement` Merck and Moderna. "Merck and Moderna Announce Phase 3 INTerpath-001 Trial of Intismeran Autogene Plus KEYTRUDA Met Endpoints of RFS and DMFS in Patients With Completely Resected Stage IIB–IV Melanoma." Joint press release, August 19, 2026. {A topline interim-analysis announcement; hazard ratios and detailed results were not disclosed with it.}

[^moderna2023]: `statement` Moderna. "Exploring One Medicine for One Patient: Individualized Neoantigen Therapies." Company blog post, 2023. {The company's own description of its pipeline; the selection algorithm's architecture and validation are not publicly disclosed.}

[^cook2014]: `paper` Cook, D. et al. "Lessons learned from the fate of AstraZeneca's drug pipeline: a five-dimensional framework." *Nature Reviews Drug Discovery*, 2014.

[^cook2014]: `paper` Cook, D. et al. "Lessons learned from the fate of AstraZeneca's drug pipeline: a five-dimensional framework." *Nature Reviews Drug Discovery*, 2014.

[^nelson2015]: `paper` Nelson, M.R. et al. "The support of human genetic evidence for approved drug indications." *Nature Genetics*, 2015. {The association is between genetic support and eventual approval; it does not show that adding genetics to a programme causes success.}

[^nelson2015]: `paper` Nelson, M.R. et al. "The support of human genetic evidence for approved drug indications." *Nature Genetics*, 2015. {The association is between genetic support and eventual approval; it does not show that adding genetics to a programme causes success.}

[^chen2019]: `paper` Chen, L. et al. "Hidden bias in the DUD-E dataset leads to misleading performance of deep learning in structure-based virtual screening." *PLoS ONE*, 2019. {The bias sits in the benchmark rather than in any one model, which makes published virtual-screening gains hard to compare.}

[^chen2019]: `paper` Chen, L. et al. "Hidden bias in the DUD-E dataset leads to misleading performance of deep learning in structure-based virtual screening." *PLoS ONE*, 2019. {The bias sits in the benchmark rather than in any one model, which makes published virtual-screening gains hard to compare.}