selekoa

A documentary history of machine intelligence.

Records found and checked by machines, not reviewed by an editor · What this means →

Propositions · 6 propositions · 30 records · 1987–2026

AI in science

Which located texts state, contest or bear on what AI does, or will do, in a science? In biology, from 1987 to 2026: whether predicted protein structures help to design drugs, whether AI models can simulate cells, whether drug candidates found with AI have entered clinical trials and pass them more often, whether machine learning makes new proteins designable, and whether biology is information that AI can read or program. For each proposition, which text is the earliest located?

The collection records AI in the sciences field by field: the institutions and programmes, the models and papers, the people credited with them and the claims made about what AI does to the field. It is not a survey of any field. Biology is the first field recorded. This page sets out the claims. Each is a proposition with one criterion, phrased generically and attributed to no one. A text that meets the criterion articulates the proposition. A text that disputes it contests it. A text that comes close bears on it, and its record names what is missing. The institutions, papers and people are recorded as dated statements of the documents that state them, on their own pages. Texts are listed in date order, and the order records no influence between them.

A machine selected and arranged these records. No human review of the selection or framing is recorded.

Download research starter

The original sources first, in date order, then every proposition on this page with its criterion, earliest located text and records. Each record gives its original source and the place in it, its role, register and dates, the qualification and the missing element of a context record, then the Selekoa identifiers. Stored paraphrases and permitted short quotes are included. Original documents are not.

Original sources for a reference manager. Every source this export cites, for Zotero, EndNote and similar tools. Cite the original documents, not Selekoa, and check each entry against them.

How we searched

The collection records what located texts say about six propositions on AI in biology. The propositions were written before the search, each with one criterion. “Earliest located” means the earliest text found by the searches below. It is not a claim that no earlier text exists.

Venues and years searched

Biology, texts from 1987 to 2026

  • Europe PMC searches by title and abstract terms, sorted by date, for structure prediction with drug design (1960–1995), “virtual cell” and whole-cell models with machine learning (1960–2023), and the other propositions’ terms; then the core records for abstracts and dates.
  • Article pages on nature.com, arXiv abstract pages and the bioRxiv records, retrieved directly.
  • Company newsrooms and releases: Sumitomo Dainippon Pharma and Exscientia, Insilico Medicine, Isomorphic Labs, EvolutionaryScale and Google DeepMind.
  • Several texts were looked up by title as candidates rather than found by the term searches. Such lookups find what is already known and favour well-cited texts.
  • Sources were retrieved directly by script on 2026-10-30.

A cross-family check, 2026-10-30

  • A model from another family answered each proposition blind from its criterion and searched for earlier texts, then compared its answers with the records.
  • It located earlier texts for two propositions, a 1990 paper on cell signalling and a 2013 company announcement of a trial, which hold those roles from the admission of 2026-10-31. For two others it located earlier candidates whose fit to the criterion is unsettled, so no earliest located text is named for them yet.
Methods
  • Every quotation was compared with the retrieved text of its own source. Paraphrases are labelled as such.
  • Abstracts read in PubMed or Europe PMC records are the authors’ own text. The publishers’ pages of several papers refused retrieval, and the check read their dates in Internet Archive captures.
  • Each record was applied against its proposition’s criterion as worded. A text that argues something adjacent bears on the proposition and its record says what is missing.
  • A fresh-context machine audit by Claude models re-read a sample of 72 records of the research pack, and a check by an OpenAI GPT model re-examined 120 items. No person has reviewed the records.

Known gaps

  • The searches for the virtual-cell, protein-design and information propositions were weakly bounded. Two 2019 preprints on protein design were read only by the check.
  • The check read the main text of the 2024 analysis of clinical-trial success rates through an author-uploaded copy relayed by a web reader; the publisher’s text and its supplementary tables were not read.
  • Not searched: clinical-trial registries, conference talks, patents and texts not in English.
  • Other sciences are not yet recorded.

A text found later that is earlier than an earliest located text becomes the new earliest located text, and the earlier headline stays on record as a correction.

How to read a record
Articulates
A passage meets the proposition’s criterion.
Contests
A passage disputes the proposition or a premise it depends on.
Bears on
A passage comes close without meeting the criterion; the record names what is missing.
Earliest located
The earliest articulation the searches found, dated by its wording. It is not a claim that no earlier text exists.
Earliest located use of the name
The earliest located text that uses the proposition’s usual name. A title can carry the name without stating the proposition.
Dates
The first date is the one the work is dated by. “Wording attested” gives the date from which the quoted wording is known, when it survives only in a later copy. “Claimed” gives an earlier talk, draft or presentation that does not date the wording.
Register: the kind of text
Fiction, satire, a dream frame or a relay of another author is labelled. A passage in fiction records what the work says, not its author’s view. Unlabelled records are nonfiction.
Limits of this page
  • “Earliest located” means the earliest text found by the searches described under “How we searched”. It is not a claim of first use or first statement.
  • A record documents what a text says. It is not evidence that the proposition is true, that any model can do what a text says, or that any drug works.
  • A company’s statement that a drug candidate entered a trial, or of how it was found, is the company’s. No trial registry was read.
  • The first criterion is not limited to machine learning: computational prediction of protein structure predates it, and the earliest located text shows this.
  • A machine selected and classified these records, and machine checks from two model families re-examined them. No person has reviewed them.

biology · Biology

Protein structure prediction and drug design, models of the cell, drug candidates found with AI and their clinical trials, the design of new proteins, and biology described as information.

P1 Predicted protein structures aid drug design

Criterion. The text states that protein structures obtained by computational prediction, rather than by experiment, are or will be of use in designing or discovering drugs.

Usual name: “structure-based drug design”

Filled dots articulate, open dots contest, small dots bear on the proposition; a large dot is an earliest located text. A line below a dot runs to the date from which its wording is attested.

Texts that articulate or contest it (3)

  1. Articulates · Earliest locatedT. L. Blundell; B. L. Sibanda; M. J. E. Sternberg; J. M. Thornton · Knowledge-based prediction of protein structures and the design of novel moleculesThe Nature review "Knowledge-based prediction of protein structures and the design of novel molecules" (Blundell, Sibanda, Sternberg and Thornton, 26 March 1987) states that knowledge-based prediction techniques are of value in predicting receptor structures to aid the design of drugs.1987-03-26 Inspect claim & evidence ↗Document page →
  2. ArticulatesDeepMind · AlphaFold — CASP14 research account (2020)DeepMind's account of AlphaFold at CASP14 (dated 30 November 2020) states that insights from protein structure predictions could enable more precise work on drug development, complementing existing experimental methods.2020-11-30 Inspect claim & evidence ↗Document page →
  3. Articulates · Earliest located use of the nameLyu J; Kapolka N; Gumpper R; Shoichet BK; Roth BL · AlphaFold2 structures guide prospective ligand discoveryLyu et al. (Science, published online 16 May 2024) report that prospective docking against unrefined AlphaFold2 models of two receptors gave hit rates similar to docking against experimental structures, and state that such models remain relevant for ligand discovery, extending the domain of structure-based drug design.2024-05-16 Inspect claim & evidence ↗Document page →
Texts that bear on it without meeting the criterion (5)
  1. Bears onSternberg MJ; Lewis RA; King RD; Muggleton S · Modelling the structure and function of enzymes by machine learningThe Faraday Discussions paper "Modelling the structure and function of enzymes by machine learning" (Sternberg, Lewis, King and Muggleton, 1992) reports applying the machine-learning program GOLEM both to protein secondary-structure prediction and, separately, to modelling a structure-activity relationship in drug design.Missing: Does not state that predicted protein structures aid drug design: the structure-prediction and drug-design applications are separate, and the drug-design model rests on ligand activities, not on a predicted protein structure.1992 Inspect claim & evidence ↗Document page →
  2. Bears onAsher Mullard · What does AlphaFold mean for drug discovery?Mullard's Nature Reviews Drug Discovery news article "What does AlphaFold mean for drug discovery?" (14 September 2021) says in its standfirst that the implications of AlphaFold and RoseTTAFold for drug discovery are more incremental than their advance in structure prediction.Missing: Assesses the degree of the implications for drug discovery ("more incremental"); it does not state that predicted structures are of no use, nor that they are of use.2021-09-14 Inspect claim & evidence ↗Document page →
  3. Bears onWong F; Krishnan A; Zheng EJ; Stärk H; Manson AL; Earl AM; Jaakkola T; Collins JJ · Benchmarking AlphaFold-enabled molecular docking predictions for antibiotic discoveryWong et al. (Molecular Systems Biology, 6 September 2022) report that docking compounds against AlphaFold2-predicted structures of essential Escherichia coli proteins gave weak performance in their benchmark, and state that advances in modelling protein-ligand interactions are needed to better harness AlphaFold2 for drug discovery.Missing: A caution about present docking methods; it neither states nor disputes that predicted structures aid drug discovery, and it frames AlphaFold2 as something to be "better harnessed".2022-09-06 Inspect claim & evidence ↗Document page →
  4. Bears onCarrie Arnold · AlphaFold touted as next big thing for drug discovery — but is it?Arnold's Nature news article (22 September 2023) says in its standfirst that questions remain about whether AlphaFold can really shake up the pharmaceutical industry.Missing: Reports an open question about degree (whether AlphaFold can "really shake up" the industry); the readable text neither states nor disputes the criterion.2023-09-22 Inspect claim & evidence ↗Document page →
  5. Bears onThomas C. Terwilliger; Dorothee Liebschner; Tristan I. Croll · AlphaFold predictions are valuable hypotheses and accelerate but do not replace experimental structure determinationTerwilliger et al. (Nature Methods, published online 30 November 2023) suggest treating AlphaFold predictions as exceptionally useful hypotheses and carrying out experimental structure determination to verify structural details, particularly interactions not included in the prediction.Missing: Addresses whether predictions can substitute for experimental structures, not whether they aid drug design.2023-11-30 Inspect claim & evidence ↗Document page →

P2 AI models can simulate or predict the behaviour of cells (a "virtual cell")

Criterion. The text states that a machine-learning or other AI model simulates, or will be able to simulate or predict, the behaviour of a living cell, such as its growth or its responses to perturbations.

Usual names: “virtual cell”, “AI Virtual Cell”

Filled dots articulate, open dots contest, small dots bear on the proposition; a large dot is an earliest located text. A line below a dot runs to the date from which its wording is attested.

Texts that articulate or contest it (5)

  1. Articulates · Earliest locatedD. Bray · Intracellular signalling as a parallel distributed processBray (Journal of Theoretical Biology, 22 March 1990) describes a procedure by which a parallel distributed processing (PDP) network could simulate a cell signalling pathway, and states that the analogy with PDP networks shows how living cells can recognise combinations of environmental influences.1990-03-22 Inspect claim & evidence ↗Document page →
  2. ArticulatesJianzhu Ma; Michael Ku Yu; Samson Fong; Keiichiro Ono; Eric Sage; Barry Demchak; Roded Sharan; Trey Ideker · Using deep learning to model the hierarchical structure and function of a cellMa et al. (Nature Methods, published online 5 March 2018) state that DCell, a "visible" neural network embedded in the hierarchical structure of the subsystems of a eukaryotic cell and trained on several million genotypes, simulates cellular growth nearly as accurately as laboratory observations.2018-03-05 Inspect claim & evidence ↗Document page →
  3. ArticulatesCharlotte Bunne; Yusuf Roohani; Yanay Rosen; Stephen R. Quake · How to Build the Virtual Cell with Artificial Intelligence: Priorities and OpportunitiesBunne et al. (arXiv v1, 18 September 2024) propose a vision of "AI-powered Virtual Cells" whose representations are learned from biological data and that facilitate in silico experiments to predict and understand cell behaviour.2024-09-18 Inspect claim & evidence ↗Document page →
  4. ArticulatesRoohani YH; Hua TJ; Tung PY; Hsu PD; Goodarzi H; Burke DP · Virtual Cell Challenge: Toward a Turing test for the virtual cellRoohani et al. (Cell, June 2025) describe virtual cells as an emerging frontier at the intersection of AI and biology, name predicting cellular responses to perturbations as a key goal of such cell state models, and announce the Arc Institute's Virtual Cell Challenge as a recurring benchmark competition toward that goal.2025-06 Inspect claim & evidence ↗Document page →
  5. ContestsConstantin Ahlmann-Eltze; Wolfgang Huber; Simon Anders · Deep-learning-based gene perturbation effect prediction does not yet outperform simple linear baselinesAhlmann-Eltze, Huber and Anders (Nature Methods, published online 4 August 2025) report that none of five foundation models and two other deep-learning models outperformed deliberately simple baselines at predicting transcriptome changes after genetic perturbations, and conclude that the goal of predicting the outcome of not-yet-performed experiments is still elusive.2025-08-04 Inspect claim & evidence ↗Document page →
Texts that bear on it without meeting the criterion (1)
  1. Bears on · Earliest located use of the nameTomita M · Whole-cell simulation: a grand challenge of the 21st centuryTomita's article "Whole-cell simulation: a grand challenge of the 21st century" (Trends in Biotechnology, June 2001) states that the complex behaviour of the cell cannot be determined or predicted unless a computer model of the cell is constructed, and reports that the E-CELL project has completed "a couple of" virtual cell models.Missing: No AI or machine-learning element: the virtual cell is a computer simulation built from biological data.2001-06 Inspect claim & evidence ↗Document page →

Headlines replaced

  • Ma and others in Nature Methods (5 March 2018), headlined as earliest located by the research pack of 2026-10-30. Earlier text found by the cross-family check: Bray’s paper of 22 March 1990, which describes how a parallel distributed processing network could simulate a cell signalling pathway. The criterion covers a cell’s responses and does not require a whole-cell model. Recorded 2026-10-31.

P3 A drug candidate discovered or designed with AI has entered human clinical trials

Criterion. The text states that a named drug candidate, described as discovered, designed or created using AI, has begun testing in humans in a clinical trial.

Filled dots articulate, open dots contest, small dots bear on the proposition; a large dot is an earliest located text. A line below a dot runs to the date from which its wording is attested.

Texts that articulate or contest it (5)

  1. Articulates · Earliest locatedBerg · Berg Launches First-Ever Stratified Medicine Clinical Trial To Investigate Cancer Metabolism Of Solid TumorsBerg’s news release of 21 November 2013 announces a phase Ib clinical trial at three named centres to study its investigational compound BPM 31510, which it describes as discovered and developed by its Interrogative Biology platform, “an artificial intelligence-based system”.2013-11-21 Inspect claim & evidence ↗Document page →
  2. ArticulatesSumitomo Pharma (then Sumitomo Dainippon Pharma) · Sumitomo Dainippon Pharma and Exscientia Joint Development New Drug Candidate Created Using Artificial Intelligence (AI) Begins Clinical StudyA news release of Sumitomo Dainippon Pharma and Exscientia (30 January 2020) announces that a phase I clinical study of DSP-1181, which it says was created using artificial intelligence, has been initiated in Japan.2020-01-30 Inspect claim & evidence ↗Document page →
  3. ArticulatesExscientia plc (filed with the US SEC) · Exscientia plc IPO prospectus (424B4)Exscientia’s IPO prospectus (30 September 2021) states that it began a Phase 1 trial of EXS21546 in December 2020 and that it “originated the first three AI-designed precision drug candidates to enter human clinical trials”.2021-09-30 Inspect claim & evidence ↗Document page →
  4. ArticulatesZuojun Xu; Feng Ren; Ping Wang; Alex Zhavoronkov · A generative AI-discovered TNIK inhibitor for idiopathic pulmonary fibrosis: a randomized phase 2a trialXu et al. (Nature Medicine, published online 3 June 2025) report a phase 2a randomized, placebo-controlled trial of rentosertib, which they describe as an AI-generated small-molecule inhibitor of a target discovered using generative AI.2025-06-03 Inspect claim & evidence ↗Document page →
  5. ArticulatesInSilico Medicine Cayman TopCo, published on HKEXnews · Insilico Medicine: Initiation of Rentosertib Phase III Clinical TrialInsilico Medicine’s announcement of 8 July 2026 states that rentosertib, which it says was discovered and developed with the support of its generative AI platform, had entered a Phase III clinical trial.2026-07-08 Inspect claim & evidence ↗Document page →
Texts that bear on it without meeting the criterion (1)
  1. Bears onKp Jayatunga M; Ayers M; Bruens L; Jayanth D; Meier C · How successful are AI-discovered drugs in clinical trials? A first analysis and emerging lessonsJayatunga et al. (Drug Discovery Today, published online 30 April 2024) state that a growing number of drugs and vaccines have been discovered using AI and analyse the clinical pipelines of what they call AI-native biotech companies.Missing: Names no specific candidate in the abstract read; the criterion requires a named candidate stated to have begun human testing.2024-04-30 Inspect claim & evidence ↗Document page →

Headlines replaced

  • The release of Sumitomo Dainippon Pharma and Exscientia of 30 January 2020, headlined as earliest located by the research pack of 2026-10-30. Earlier text found by the cross-family check: Berg’s release of 21 November 2013, which announces a trial of a compound it describes as discovered with an AI-based platform. It is the company’s own statement. Recorded 2026-10-31.

P6 AI-discovered drug candidates succeed in clinical trials more often than historical averages

Criterion. The text states that drug candidates discovered or designed with AI have passed at least one phase of clinical trials at a higher rate than historical industry averages.

Filled dots articulate, open dots contest, small dots bear on the proposition; a large dot is an earliest located text. A line below a dot runs to the date from which its wording is attested.

Texts that articulate or contest it (1)

  1. Articulates · Earliest locatedKp Jayatunga M; Ayers M; Bruens L; Jayanth D; Meier C · How successful are AI-discovered drugs in clinical trials? A first analysis and emerging lessonsJayatunga et al. (Drug Discovery Today, published online 30 April 2024) state that, in their analysis, AI-discovered molecules have an 80–90% success rate in phase I, substantially higher than historic industry averages, and a phase II success rate of about 40%, comparable to those averages, on a limited sample.2024-04-30 Inspect claim & evidence ↗Document page →
Texts that bear on it without meeting the criterion (1)
  1. Bears onZuojun Xu; Feng Ren; Ping Wang; Alex Zhavoronkov · A generative AI-discovered TNIK inhibitor for idiopathic pulmonary fibrosis: a randomized phase 2a trialXu et al. (Nature Medicine, published online 3 June 2025) state that AI-discovered drugs have experienced similar levels of phase 2 failure as non-AI-discovered drugs, and that none has so far progressed through phase 3.Missing: Addresses phases 2 and 3 only; it neither states nor disputes a higher success rate in phase I.2025-06-03 Inspect claim & evidence ↗Document page →

P4 Machine-learning methods make proteins with new functions designable

Criterion. The text states that machine-learning models can be used, or will contribute, to designing proteins with functions, or sequences, not found among natural proteins.

Earliest located
Not assigned. Withheld until the criterion is decided (maintainer decision of 2026-10-30). The cross-family check located earlier candidates, Schneider and Wrede (1994) on sequence design assisted by neural networks and Greener, Moffat and Jones (arXiv, 26 June 2018) on sequences for a novel protein topology. Whether either meets the criterion’s “not found among natural proteins” is unsettled.
Filled dots articulate, open dots contest, small dots bear on the proposition; a large dot is an earliest located text. A line below a dot runs to the date from which its wording is attested.

Texts that articulate or contest it (3)

  1. ArticulatesAnishchenko I; Chidyausiku TM; Ovchinnikov S; Pellock SJ; Baker D · De novo protein design by deep network hallucinationThe bioRxiv preprint "De novo protein design by deep network hallucination" (Anishchenko, Chidyausiku, Ovchinnikov, Pellock and Baker, v1 of 23 July 2020) states that deep networks trained to predict native protein structures can be inverted to design new proteins and should contribute to the de novo design of proteins with new functions.2020-07-23 Inspect claim & evidence ↗Document page →
  2. ArticulatesJoseph L. Watson; David Juergens; Nathaniel R. Bennett; Brian L. Trippe; Jason Yim; Helen E. Eisenach; Woody Ahern; Andrew J. Borst; Robert J. Ragotte; Lukas F. Milles; Basile I. M. Wicky; Nikita Hanikel; Samuel J. Pellock; Alexis Courbet; William Sheffler; Jue Wang; Preetham Venkatesh; Isaac Sappington; Susana Vázquez Torres; Anna Lauko; Valentin De Bortoli; Emile Mathieu; Sergey Ovchinnikov; Regina Barzilay; Tommi S. Jaakkola; Frank DiMaio; Minkyung Baek; David Baker · De novo design of protein structure and function with RFdiffusionWatson et al. (Nature, published online 11 July 2023) state that their deep-learning method RFdiffusion enables the design of diverse functional proteins from simple molecular specifications.2023-07-11 Inspect claim & evidence ↗Document page →
  3. ArticulatesEvolutionaryScale · ESM3: Simulating 500 million years of evolution with a language modelEvolutionaryScale's ESM3 announcement (dated 25 June 2024) states that its language model generated a new fluorescent protein, esmGFP, with a sequence only 58% similar to the closest known fluorescent protein.2024-06-25 Inspect claim & evidence ↗Document page →
Texts that bear on it without meeting the criterion (2)
  1. Bears onPo-Ssu Huang; Scott E. Boyken; David Baker · The coming of age of de novo protein designHuang, Boyken and Baker (Nature, published online 14 September 2016) state that computational methodology now allows a wide range of protein structures to be designed from scratch with atomic-level accuracy, and that it should now be possible to design new functional proteins from the ground up.Missing: No machine-learning element: the design methods are physics-based.2016-09-14 Inspect claim & evidence ↗Document page →
  2. Bears onJingxue Wang; Huali Cao; John Z. H. Zhang; Yifei Qi · Computational Protein Design with Deep Learning Neural NetworksWang, Cao, Zhang and Qi (Scientific Reports, 20 April 2018) report applying a deep-learning neural network to computational protein design and state that their results may benefit further development of computational protein design methods.Missing: Applies machine learning to protein design but does not state that it makes proteins with new functions or unnatural sequences designable.2018-04-20 Inspect claim & evidence ↗Document page →

Headlines replaced

  • The preprint of 23 July 2020 on protein design by network hallucination, headlined as earliest located by the research pack of 2026-10-30. The cross-family check located earlier candidates, from 1994 and 2018, whose fit to the criterion is unsettled. No text holds the role until the criterion is decided. Recorded 2026-10-31.

P5 Biology is an information system that AI can read, describe or program

Criterion. The text states that biology is, or can be treated as, an information system, language or code, and that AI is, or may become, the means of reading, describing or programming it.

Usual names: “digital biology”, “programmable biology”

Earliest located
Not assigned. Withheld until the criterion is decided (maintainer decision of 2026-10-30). The cross-family check located an earlier candidate, Searls (AAAI-88 proceedings, 1988), which treats genetic information as a knowledge representation for symbolic AI. Whether that meets the criterion’s AI clause is unsettled.
Earliest located use of the name
No located text uses the name.
Filled dots articulate, open dots contest, small dots bear on the proposition; a large dot is an earliest located text. A line below a dot runs to the date from which its wording is attested.

Texts that articulate or contest it (2)

  1. ArticulatesIsomorphic Labs · Introducing Isomorphic LabsThe announcement “Introducing Isomorphic Labs” (November 2021) states that biology can be thought of as an information processing system and that, just as mathematics turned out to be the right description language for physics, biology may turn out to be the perfect type of regime for the application of AI.2021-11 · wording attested 2026-10-30 Inspect claim & evidence ↗Document page →
  2. ArticulatesEvolutionaryScale · ESM3: Simulating 500 million years of evolution with a language modelEvolutionaryScale's ESM3 announcement (dated 25 June 2024) states that biology is written in a language not yet understood, and that learning to read and write in the code of life would make biology programmable.2024-06-25 Inspect claim & evidence ↗Document page →
Texts that bear on it without meeting the criterion (1)
  1. Bears onLeroy Hood; David Galas · The digital code of DNAHood and Galas (Nature, 23 January 2003) state that the discovery of the structure of DNA engendered a new view of biology as an information science.Missing: No AI element: the information view is tied to DNA and genomics, not to AI as the means of reading or describing biology.2003-01-23 Inspect claim & evidence ↗Document page →

Headlines replaced

  • The Isomorphic Labs launch article of November 2021, headlined as earliest located by the research pack of 2026-10-30. The cross-family check located an earlier candidate, a 1988 conference paper that treats genetic information as a knowledge representation for symbolic AI, whose fit to the criterion is unsettled. No text holds the role until the criterion is decided. Recorded 2026-10-31.