CoChat
Sign Up
Cited by 62Computational Drug Discovery MethodsMachine Learning in Materials ScienceArtificial Intelligence in Healthcare and Education

Artificial Intelligence in Small-Molecule Drug Discovery: A Critical Review of Methods, Applications, and Real-World Outcomes

Open access
Niazi, Sarfaraz K.·2025·Pharmaceuticals
DOI 10.3390/ph18091271PMC PMC12472608
Open this paper in CoChat — free accountOpen PDF

Abstract

Artificial intelligence (AI) is emerging as a valuable complementary tool in small-molecule drug discovery, augmenting traditional methodologies rather than replacing them. This review examines the evolution of AI from early rule-based systems to advanced deep learning, generative models, diffusion models, and autonomous agentic AI systems, highlighting their applications in target identification, hit discovery, lead optimization, and safety prediction. We present both successes and failures to provide a balanced perspective. Notable achievements include baricitinib (BenevolentAI/Eli Lilly, an existing drug repurposed through AI-assisted analysis for COVID-19 and rheumatoid arthritis), halicin (MIT, preclinical antibiotic), DSP-1181 (Exscientia, discontinued after Phase I), and ISM001-055/rentosertib (Insilico Medicine, positive Phase IIa results). However, several AI-assisted compounds have also faced challenges in clinical development. DSP-1181 was discontinued after Phase I, despite a favorable safety profile, highlighting that the acceleration of discovery timelines by AI does not guarantee clinical success. Despite progress, challenges such as data quality, model interpretability, regulatory hurdles, and ethical concerns persist. We provide practical insights for integrating AI into drug discovery workflows, emphasizing hybrid human-AI approaches and the emergence of agentic AI systems that can autonomously navigate discovery pipelines. A critical evaluation of current limitations and future opportunities reveals that while AI offers significant potential as a complementary technology, realistic expectations and careful implementation are crucial for delivering innovative therapeutics.

The claims and statistics of this paper haven’t been checked yet.

Run the check — free, no account needed. It reads the open-access text and checks the sources it cites.

A few free checks a day · papers already verified open instantly · results are public

Integrity signals

Registry

Crossref · 10.3390/ph18091271

Cites

125 of 132 references verified0 citation errors · 0 not checkable

Statistics

Claims the paper makes

The claims and statistics of this paper haven’t been checked yet.

Run the check — free, no account needed. It reads the open-access text and checks the sources it cites.

A few free checks a day · papers already verified open instantly · results are public

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

  • Protocol automation can be reformulated as retrieval-based procedural synthesis tasks, and this approach reduces setup times by orders of magnitude in reported cases while improving standardization across laboratories.

    “This approach shifts the work from manual model modification to structured supervision, reducing setup times by orders of magnitude in reported cases while improving standardization across laboratories [45].”

    From The next paradigm in bioinformatics: a review of multi-agent systems and foundational models for end-to-end scientific discovery (2026) · cites this paper as [45]

    Section 6.5 reports that AI-integrated automated synthesis can shorten small-molecule discovery timelines, but says the extent varies by application. It does not describe retrieval-based procedural synthesis, structured supervision, or setup-time reductions by orders of magnitude. Section 9.2 identifies standardization as a need, not an improvement demonstrated across laboratories by this approach. Page numbers were not provided in our copy. (LLM, conf=0.82)

    Experimental — accuracy not yet measured

  • Generative modeling frameworks are increasingly used to propose individualized therapeutic hypotheses or optimized drug combinations in the presence of explicit clinical constraints.

    “Generative modeling frameworks, which often integrate sequence and graph transformers with latent variable representations, are increasingly used to propose individualized therapeutic hypotheses or optimized drug combinations in the presence of explicit clinical constraints [45, 46].”

    From The next paradigm in bioinformatics: a review of multi-agent systems and foundational models for end-to-end scientific discovery (2026) · cites this paper as [45]

    Section 3.4 (page not provided) describes generative models that design molecules with desired properties and can incorporate property constraints. It does not describe their use to propose individualized therapeutic hypotheses or optimized drug combinations under explicit clinical constraints; the constraints discussed are molecular, not clinical. (LLM, conf=0.86)

    Experimental — accuracy not yet measured

  • Reported implementations suggest a substantial increase in productivity in screening and triage cycles when reasoning and execution levels are closely correlated.

    “Reported implementations suggest a substantial increase in productivity in screening and triage cycles when reasoning and execution levels are closely correlated [45, 47].”

    From The next paradigm in bioinformatics: a review of multi-agent systems and foundational models for end-to-end scientific discovery (2026) · cites this paper as [45]

    Section 5.3 reports more effective screening triage: the AI-ranked subset had fivefold enrichment in confirmed antiviral activity, and two compounds reached preclinical studies in eight months rather than the typical 18–24 months. The cited passages do not establish that these gains occurred when “reasoning and execution levels” were closely correlated, so that condition is unsupported. (LLM, conf=0.85)

    Experimental — accuracy not yet measured

  • By explicitly tracing the provenance of evidence and ranking candidate hypotheses, such systems can reduce the time required for structured review from weeks to hours, while preserving verifiability and reproducibility.

    “By explicitly tracing the provenance of evidence and ranking candidate hypotheses, such systems can reduce the time required for structured review from weeks to hours, while preserving verifiability and reproducibility.”

    From The next paradigm in bioinformatics: a review of multi-agent systems and foundational models for end-to-end scientific discovery (2026) · cites this paper as [45]

    Section 6.4 describes agentic AI systems that synthesize scientific literature and generate hypotheses, supporting part of the description. It does not say they trace evidence provenance, rank hypotheses, or reduce structured-review time from weeks to hours. Section 8 discusses reproducibility of model predictions, not whether these review systems preserve verifiability and reproducibility; no page numbers are available in our copy. (LLM, conf=0.82)

    Experimental — accuracy not yet measured

  • The value of agentic AI in drug discovery and personalized medicine emerges when reasoning, action, and evidence monitoring operate as an integrated system capable of accelerating insight without compromising reproducibility or governance.

    “As recent implementations indicate, the value of agentic AI in drug discovery and personalized medicine emerges when reasoning, action, and evidence monitoring operate as an integrated system capable of accelerating insight without compromising reproducibility or governance [27, 33, 34, 45–48].”

    From The next paradigm in bioinformatics: a review of multi-agent systems and foundational models for end-to-end scientific discovery (2026) · cites this paper as [45]

    Section 6.4 describes agentic systems that combine capabilities to propose experiments, interpret results, and refine hypotheses, and reports early lead-discovery implementations with reduced human intervention. It also says these systems still require human oversight. Section 8 treats reproducibility and continuous monitoring as validation requirements, but the paper does not establish that integrated agentic systems accelerate insight without compromising them, or extend its agentic-AI findings to personalized medicine. (LLM, conf=0.86)

    Experimental — accuracy not yet measured

Self-check

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

The self-check needs the paper’s full text.

Cites

Flagged first, then by citations

Research works printed but not linked: 16

  1. [32]Hoogeboom E. Satorras V.G. Vignac C. Welling M. Equivariant Diffusion for Molecule Generation in 3D Proceedings of the 39th International Conference on Machine Learning Baltimore, MD, USA 17–23 July 2022 PMLR Cambridge, MA, USA 2022 8867 8887 Available online: https://proceedings.mlr.press/v162/hoogeboom22a.html (accessed on 3 August 2025)
  2. [51]Wallach I. Dzamba M. Heifets A. AtomNet: A Deep Convolutional Neural Network for Bioactivity Prediction in Structure-Based Drug Discovery arXiv 2015 1510.02855 Available online: https://arxiv.org/abs/1510.02855 (accessed on 3 August 2025) · arXiv
  3. [55]Baker M. Bode K. Cheetham T. Del Carratore F. Text mining in drug discovery: Applications and opportunities Drug Discov. Today 2017 22 271 279 10.1016/j.drudis.2016.11.011 · DOI
  4. [61]Parks C.D. Gaieb Z. Chiu M. Yang H. Shao C. Walters W.P. Deep learning models for virtual screening: A critical review Curr. Opin. Struct. Biol. 2020 63 82 92 10.1016/j.sbi.2020.04.001 · DOI
  5. [64]Gilmer J. Schoenholz S.S. Riley P.F. Vinyals O. Dahl G.E. Neural Message Passing for Quantum Chemistry Proceedings of the 34th International Conference on Machine Learning Sydney, Australia 6–11 August 2017 PMLR: Cambridge, MA, USA, 2017 1263 1272 Available online: https://proceedings.mlr.press/v70/gilmer17a.html (accessed on 3 August 2025)
  6. [65]Hu W. Liu B. Gomes J. Zitnik M. Liang P. Pande V. Strategies for Pre-Training Graph Neural Networks Proceedings of the International Conference on Learning Representations Addis Ababa, Ethiopia 30 April 2020 Available online: https://openreview.net/forum?id=HJlWWJSFDH (accessed on 3 August 2025)
  7. [78]Kingma D.P. Welling M. Auto-encoding Variational Bayes arXiv 2013 1312.6114 Available online: https://arxiv.org/abs/1312.6114 (accessed on 3 August 2025) · arXiv
  8. [79]Guimaraes G.L. Sanchez-Lengeling B. Outeiral C. Farias P.L.C. Aspuru-Guzik A. Objective-Reinforced Generative Adversarial Networks (ORGAN) for Sequence Generation Models arXiv 2017 1705.10843 Available online: https://arxiv.org/abs/1705.10843 (accessed on 3 August 2025) · arXiv
  9. [89]Holzinger A. Biemann C. Pattichis C.S. Kell D.B. What do We Need to Build Explainable AI Systems for the Medical Domain? arXiv 2017 1712.09923 Available online: https://arxiv.org/abs/1712.09923 (accessed on 3 August 2025) · arXiv
  10. [91]Bahdanau D. Cho K. Bengio Y. Neural Machine Translation by Jointly Learning to Align and Translate arXiv 2014 1409.0473 Available online: https://arxiv.org/abs/1409.0473 (accessed on 3 August 2025) · arXiv
  11. [103]Hopkins A.L. Exscientia progress report: AI-designed molecules in the clinic Drug Discov. Today 2022 27 1903 1905 10.1016/j.drudis.2022.05.015 · DOI
  12. [122]Hinton G. Vinyals O. Dean J. Distilling the knowledge in a neural network arXiv 2015 1503.02531 Available online: https://arxiv.org/abs/1503.02531 (accessed on 3 August 2025) · arXiv
  13. [130]Huang K. Fu T. Gao W. Zhao Y. Roohani Y. Leskovec J. Therapeutics Data Commons: Machine learning datasets and tasks for drug discovery and development arXiv 2021 2102.09548 Available online: https://arxiv.org/abs/2102.09548 (accessed on 3 August 2025) · arXiv
  14. [132]Sheridan R.P. Time to change how we do QSAR J. Comput. Aided Mol. Des. 2013 27 847 855 10.1007/s10822-013-9678-y · DOI
  15. [141]Reardon S. Rise of robot chemists sparks debate over AI patents Nature 2019 574 155 156 10.1038/d41586-019-02999-2 · DOI
  16. [145]Rashidi P. Bihorac A. Rashidi H. Johnson J.A. Tighe P.J. Digital twins: The convergence of physical and virtual in healthcare Nat. Biotechnol. 2019 37 1395 1397 10.1038/s41587-019-0339-0 · DOI

Cited by

Sorted by citations

Related

Authors
Niazi, S. K.
Affiliations
University of Illinois Chicago (US)
Corresponding author
Niazi, Sarfaraz K.
ORCID
Niazi, Sarfaraz K.
Venue
Pharmaceuticals
Publisher
MDPI AG
Published
2025-08-26
Volume
18
Issue
9
Pages
1271
ISSN
1424-8247 (ISSN-L)
DOI
10.3390/ph18091271
PMC
PMC12472608
OpenAlex
W4413659714
Cited by
61
References
124
Citations by year
2026: 58 · 2025: 3
FWCI
3.76
Type
journal article
Language
en
Open access
Licence
Locations
  • Pharmaceuticals (publishedVersion, cc-by)
  • Europe PMC (PubMed Central) (publishedVersion, cc-by)
  • PubMed (publishedVersion)
  • PubMed Central (publishedVersion, cc-by)
  • DOAJ (DOAJ: Directory of Open Access Journals) (submittedVersion, cc-by-sa)
Topics
  • Physical Sciences › Computer Science › Computational Theory and Mathematics › 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, small molecule drug discovery, deep learning, generative models, diffusion models, agentic AI, target identification, hit identification, lead optimization, drug safety prediction, drug repurposing, baricitinib, model interpretability, AI ethics and regulation

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.

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