CoChat

The DOI matches its Crossref record; 197 of the 197 references it cites were found in a registry or an index; 55 of the 90 sources its claims cite were checked for support.

Artificial Intelligence for Drug Discovery: Are We There Yet?

Open access
Hasselgren, Catrin, Oprea, Tudor I. · 2024 · Annual Review of Pharmacology and Toxicology

Abstract

Drug discovery is adapting to novel technologies such as data science, informatics, and artificial intelligence (AI) to accelerate effective treatment development while reducing costs and animal experiments. AI is transforming drug discovery, as indicated by increasing interest from investors, industrial and academic scientists, and legislators. Successful drug discovery requires optimizing properties related to pharmacodynamics, pharmacokinetics, and clinical outcomes. This review discusses the use of AI in the three pillars of drug discovery: diseases, targets, and therapeutic modalities, with a focus on small-molecule drugs. AI technologies, such as generative chemistry, machine learning, and multiproperty optimization, have enabled several compounds to enter clinical trials. The scientific community must carefully vet known information to address the reproducibility crisis. The full potential of AI in drug discovery can only be realized with sufficient ground truth and appropriate human intervention at later pipeline stages.

Overview

CoChat Summary

TL;DR

This review surveys AI for drug discovery across diseases, targets, and therapeutic modalities, focusing on small molecules, and finds AI has enabled compounds to enter clinical trials but no approved medicine can yet be attributed to AI. It concludes that AI4DD is not fully automated and that realizing its potential requires ground truth, reproducibility, and human intervention.

Key findings

  • AI4DD publications rose from 49 in 2011 to 333 in 2020; AI-designed compounds have entered trials, including DSP-1181 (phase I, Jan 2020), DSP-0039 (phase I, May 2021), ISM01-055/ISM018-055 (phase II, June 2023), and RLY-1971 (phase I).
  • KG-based XGBoost models: the AD model used 53 positive and 3,952 negative genes and identified five previously unassociated immunity genes (FRRS1, CTRAM, SCGB3A1, CIBAR2, TMEFF2); the ATG model used 103 positives and 3,468 negatives, with 43 of the top 251 already in the Autophagy Database, 15 with potential ATG associations, and 193 (77%) “ATG dark genes”; 7 of the top 20 and 2 bottom-predicted genes were validated by literature.
  • Generative chemistry: GENTRL produced 30,000 compounds, filtered to six novel DDR1/DDR2 inhibitor candidates; two were potent nM inhibitors, and the process from DDR1 nomination to Compound 6 took 35 days; Compound 1 was orally bioavailable in mice with a 3.5-hour half-life after a 15 mg/kg dose.
  • Challenges: no regulatory-approved medicine can be attributed to AI; ML models are often trained on tens of thousands of compounds versus 30 billion virtual compounds; Bayer and Amgen reported 33% and 11% reproducibility rates; data leakage was detected in 15 of 20 biology/medicine subject areas.

Methods

  • Design: narrative review of AI for drug discovery (AI4DD), focused primarily on small molecule drugs, organized around three pillars—diseases, targets, and therapeutic modalities—plus regulatory science and caveats.
  • Data/sample: cited literature, public databases/resources (e.g., DrugCentral, Pharos, Therapeutic Data Commons, OCHEM), and case studies of AI-designed compounds; no new primary experimental data are reported.
  • Analysis: qualitative synthesis and discussion of cited examples; no systematic search strategy, meta-analysis, or formal quantitative synthesis is stated.

Limitations

  • Paper-stated: no approved medicine can be attributed to AI, and no fully automated AI-driven drug discovery system exists; most successes rely on computational platforms supporting human decision-making.
  • Paper-stated: ML models face applicability domain, external predictivity validation, and chemical diversity concerns; KG-based ML lacks ground truth and can suffer data leakage; reproducibility crisis affects both experiments and ML.
  • Evident from text: as a narrative review, it does not report a systematic search or quantitative synthesis, and its examples are illustrative rather than exhaustive.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

Crossref · 10.1146/annurev-pharmtox-040323-040828

Cites

197 of 197 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Figures

Figure 1 Knowledge Graph representation of phenotype-protein associations. Nodes (main entities) are related by edges (arrows). Centered on the main node, 'Target', nodes listed clockwise are as follows: Tissue; Cell_type; Panther, Protein Analysis Through Evolutionary Relationships, a protein classification scheme(57); Reactome, a pathways database (58); Jensen DISEASES (19); OMIM, Online Mendelian Inheritance in Man database(59); HPO, human phenotype ontology (60); Homolog (which lists proteins with common evolutionary origin), further linked to IMPC (International Mouse Phenotype Consortium) mouse phenotype database (61); CCLE, cancer cell line encyclopedia (62); Transcript, Protein and unreviewed_protein - various types of UniProt (63) entries; GO, Gene Ontology (64) consortium terms; Ramilowski_subcellular_location, a specific experimental resource for subcellular protein localization (65). Nodes are associated to 'Target' through a variety of arrows. GO tems, for example, feature several protein properties, whereas STRING (20) loops protein protein interactions (PPIs) for 'Targets'; other resources such as HPA, the Human Protein Atlas (66), informs a variety of nodes regarding tissue, cell type and subcellular location and expression.
Figure 1 Knowledge Graph representation of phenotype-protein associations. Nodes (main entities) are related by edges (arrows). Centered on the main node, 'Target', nodes listed clockwise are as follows: Tissue; Cell_type; Panther, Protein Analysis Through Evolutionary Relationships, a protein classification scheme(57); Reactome, a pathways database (58); Jensen DISEASES (19); OMIM, Online Mendelian Inheritance in Man database(59); HPO, human phenotype ontology (60); Homolog (which lists proteins with common evolutionary origin), further linked to IMPC (International Mouse Phenotype Consortium) mouse phenotype database (61); CCLE, cancer cell line encyclopedia (62); Transcript, Protein and unreviewed_protein - various types of UniProt (63) entries; GO, Gene Ontology (64) consortium terms; Ramilowski_subcellular_location, a specific experimental resource for subcellular protein localization (65). Nodes are associated to 'Target' through a variety of arrows. GO tems, for example, feature several protein properties, whereas STRING (20) loops protein protein interactions (PPIs) for 'Targets'; other resources such as HPA, the Human Protein Atlas (66), informs a variety of nodes regarding tissue, cell type and subcellular location and expression.From the PDF
Figure 2 Chemical structure of two epithelial discoidin domain-containing receptor inhibitors (compounds 1 and 2), developed using the GENTRL (118) generative chemistry approach. Their chemical similarity with ponatinib (bottom left) is discussed in the text. The structure and bioactivity of MMV1803022, one of the community-driven anti-malarial AI competition compounds, is shown bottom center. Pfal summarizes the in vitro growth inhibition of asexual blood-stage P. falciparum (3D7) parasite maturation; Pf ATP4 indicates confirmed on-target biochemical (cytosolic [Na+] increase) inhibition. Halicin is on the bottom right (see text for details). Bioactivities were extracted from (118) for Compounds 1 and 2, from DrugCentral (47) for ponatinib, and from (119) for the anti-malarial compound. CIDs are PubChem compound identifiers.
Figure 2 Chemical structure of two epithelial discoidin domain-containing receptor inhibitors (compounds 1 and 2), developed using the GENTRL (118) generative chemistry approach. Their chemical similarity with ponatinib (bottom left) is discussed in the text. The structure and bioactivity of MMV1803022, one of the community-driven anti-malarial AI competition compounds, is shown bottom center. Pfal summarizes the in vitro growth inhibition of asexual blood-stage P. falciparum (3D7) parasite maturation; Pf ATP4 indicates confirmed on-target biochemical (cytosolic [Na+] increase) inhibition. Halicin is on the bottom right (see text for details). Bioactivities were extracted from (118) for Compounds 1 and 2, from DrugCentral (47) for ponatinib, and from (119) for the anti-malarial compound. CIDs are PubChem compound identifiers.From the PDF
Image 3
From the PDF
Figure 4 Chemical structures of RLY-1971 (migoprotafib or GDC-1971) and RLY-4008 (lirafugratinib), two compounds progressed into phase I clinical studies by Relay Therapeutics. While the exact bioactivity of RLY-4008 has not been disclosed, it lacks potency against FGFR1 and FGFR4 (163), which in turn may eliminate unwanted side-effects. CIDs are PubChem compound identifiers.
Figure 4 Chemical structures of RLY-1971 (migoprotafib or GDC-1971) and RLY-4008 (lirafugratinib), two compounds progressed into phase I clinical studies by Relay Therapeutics. While the exact bioactivity of RLY-4008 has not been disclosed, it lacks potency against FGFR1 and FGFR4 (163), which in turn may eliminate unwanted side-effects. CIDs are PubChem compound identifiers.From the PDF

Verification

Claims the paper makes

19 Valid · 24 Not valid · 24 Not checked · 20 Own results and interpretations

34 more claims · 17 with their cited sources checked

Create a free account to see all 87 claims and to verify the 3 references still open.

See all claims — free account
  • Probably not a substantial claim. 36
  • It may state two separate claims as one. 20

Create a free account to see these claims.

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

Each contradiction between two passages listed here was confirmed by an independent check.

1 of 1

Cited works that were retracted

Works this paper cites that were later retracted, from the retraction notices publishers deposit. Major when the notice came out before this paper was published.

1 of 1

Citations

Cites

1 retracted

Show all 167 references

Research works printed but not linked: 24

  1. Harris LA., 2021, Artificial intelligence: background, selected issues, and policy considerations
  2. Borthwick A, 1998, Proceedings of the Sixth Workshop on Very Large Corpora
  3. Bartusiak M., 1981, Discover Magazine, 47
  4. Bylinsky G., 1981, Fortune Magazine, 106
  5. Jassal B, 2020, Nucleic Acids Res., 48, D498
  6. Hansch C, 1995, Exploring QSAR: Fundamentals and Applications in Chemistry and Biology
  7. Natl. Cancer Inst, 2023, NCI-60 growth inhibition data
  8. Deb K, 2006, Int. J. Comput. Intell. Res., 2, 273
  9. Landrum G., 2022, RDKit.
  10. Pedregosa F, 2011, J. Mach. Learn. Res., 12, 2825
  11. Int. Counc. Harmon. Tech. Requir. Pharm. Hum. Use (ICH), 2023, M7: assessment and control of DNA reactive (mutagenic) impurities in pharmaceuticals to limit potential carcinogenic risk
  12. EMA (Eur. Med. Agency), 2018, Reflection paper on the qualification of non-genotoxic impurities
  13. Hines PA, 2022, Nat. Rev. Drug Discov.
  14. FDA (US Food Drug Admin.), 2023, Using artificial intelligence and machine learning in the development of drug and biological products
  15. FDA (US Food Drug Admin.), 2023, Artifical intelligence in drug manufacturing
  16. Exscientia, 2020, DSP-1181
  17. Yoshinaga H, 2020, 2,6-Disubstituted pyridine derivative
  18. Wills T., 2022, CAS
  19. Yoshinaga H, 2020, Condensed lactam derivative
  20. Fultinavičiūtė U., 2023, Clinical Trials Arena
  21. Aliper A, 2020, Methods of inhibiting kinases
  22. InSilico Med, 2023, Hong Kong Stock Exchange application proof of InSilico Medicine Cayman TopCo
  23. Relay Ther, 2020, Relay Therapeutics® announces a worldwide license and collaboration agreement with Genentech for RLY-1971
  24. OpenAI, 2023, GPT-4 technical report

Cited by

See all 236

Metadata

Authors
Hasselgren, C., & Oprea, T. I.
Affiliations
  • Genentech (US)
  • University of New Mexico (US)
Venue
Annual Review of Pharmacology and Toxicology
Published
2024-01-23
Volume
64
Issue
1
Pages
527-550
ISSN
  • 0362-1642 (ISSN-L)
  • 1545-4304
OpenAlex
W4384389969
Cited by
236
References
167
Citations by year
2026: 67 · 2025: 107 · 2024: 57
FWCI
3.99
Type
journal article
Language
en
Open access
Locations
  • The Annual Review of Pharmacology and Toxicology (publishedVersion, public-domain)
  • PubMed (publishedVersion)
  • arXiv (Cornell University) (submittedVersion, cc-by-nc-sa)
  • arXiv (Cornell University)
Funders
  • National Institutes of Health (U24 224370)
  • Genentech
  • National Cancer Institute (U24 224370)
Topics
  • Life Sciences › Pharmacology, Toxicology and Pharmaceutics › Drug Discovery › Computational Drug Discovery Methods
  • Physical Sciences › Materials Science › Materials Chemistry › Machine Learning in Materials Science
  • Health Sciences › Medicine › Health Informatics › Artificial Intelligence in Healthcare and Education
Keywords
artificial intelligence in drug discovery, generative chemistry, machine learning, small molecule drugs, therapeutic target, therapeutic modalities, replication crisis, clinical trial candidates, pharmacokinetics, pharmacodynamics
MeSH terms
Animals, Artificial Intelligence, Humans, Physicians, Technology, Reproducibility of Results, Drug Discovery

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

The claim verdicts this report counts are experimental: their accuracy is not yet measured.

A language model drew the claims from the paper on 2026-10-10, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-10 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (gpt-6-sol).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: Artificial Intelligence for Drug Discovery: Are We There Yet? https://app.cochat.ai/papers/10.1146/annurev-pharmtox-040323-040828
The paper
Hasselgren, C., & Oprea, T. I. (2024). Artificial Intelligence for Drug Discovery: Are We There Yet? Annual Review of Pharmacology and Toxicology. https://doi.org/10.1146/annurev-pharmtox-040323-040828

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.