Source-linked AI summary
Artificial Intelligence for Drug Discovery: Are We There Yet?
Catrin Hasselgren, Tudor I. Oprea
TL;DR
Drug discovery must optimize pharmacodynamic, pharmacokinetic, and clinical properties while navigating heterogeneous data and a complex development process. This review examines AI across diseases, targets, and therapeutic modalities, focusing on small molecules and computational platforms that support discovery and optimization. AI-designed compounds have entered clinical trials, but no medicines approved by regulatory agencies are attributed to AI in the same way as standalone AI achievements, and comprehensive automation remains unrealized.
Problem
Drug discovery requires concurrent optimization across pharmacodynamics, pharmacokinetics, and clinical outcomes, while AI systems face heterogeneous data and drug quality depends partly on human and regulatory approval.
Method
The review examines AI4DD across diseases, targets, and therapeutic modalities, emphasizing small-molecule drugs and computational platforms for target identification, prediction, generative chemistry, and multi-property optimization.
Results
AI-designed compounds have entered clinical trials, including DSP-1181 in phase I, ISM018-055 potentially in phase II, and scFPM-guided treatment associated with improved outcomes in a prospective study.
Takeaways & Limitations
AI currently supports human decision-making in drug discovery, while its potential depends on sufficient ground truth, model validation, chemical diversity, and human expertise.
Takeaways & Limitations
No medicines approved by regulatory agencies can yet be attributed to AI in the same way as standalone AI achievements, and drug discovery lacks a comprehensive AI system.
Abstract
from arXiv · showhide
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 multi-property 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.
3. Internal Medicine, UNM Health Sciences Center, Albuquerque, NM 87131
This block contains keywords, citation instructions, bibliographic details, and a DOI link for the paper.
- The listed keywords include deep learning, explainable AI, generative chemistry, knowledge graphs, machine learning, and multi-property optimization.
- The block instructs readers how to cite the paper.
- The bibliographic metadata identifies Hasselgren and Oprea, the 2024 publication year, and Annual Reviews in Pharmacology and Toxicology.
- A DOI link is supplied for the paper after publication.
1. Introduction
Drug discovery is a complex, iterative process that is incorporating AI and related technologies to improve efficiency, reduce costs and animal experiments, and accelerate treatment development. This review examines AI4DD across diseases, targets, and therapeutic modalities, primarily for small-molecule drugs, while also addressing regulatory science and caveats.
- Drug discovery is an iterative, multifaceted process spanning target identification, optimization, clinical development, regulatory review, and medical deployment.The pipeline includes interdependent scientific, clinical, regulatory, and post-marketing activities.
- Data science, informatics, and AI are being incorporated to improve efficiency, reduce costs and animal experiments, and accelerate novel treatment development.
- Successful approval requires concurrent optimization of pharmacodynamic, pharmacokinetic, and clinical-outcome properties.These include efficacy and drug-target interactions, ADMET and safety, therapeutic intent, and adverse outcomes.
- The review examines AI4DD across diseases, targets, and therapeutic modalities, primarily focusing on small-molecule drugs.It also includes regulatory science and selected caveats of AI4DD.
2. Emergence of artificial intelligence for drug discovery
AI for drug discovery has expanded from early computational models into a central research and commercial area, but heterogeneous data and the human-mediated nature of drug quality remain important challenges. The section traces this emergence and its implications for machine-learning-based drug-likeness assessment.
- The knowledge deficit: Drug discovery AI must integrate heterogeneous, variable-quality data and connect evidence through intelligent reasoning and pattern recognition.
- An incomplete history of AI for small molecule drug design: Early drug-design efforts included computer-aided drug design initiatives in 1981 and the founding of the Journal of Computer-Aided Drug Design in 1987.
- 1998 machine-learning models demonstrated that chemical features could distinguish compounds proposed for biological testing from compounds lacking pharmaceutical use.
- Drug quality is not intrinsic to chemicals because medicinal use depends on regulatory and human approval, which can later be revised or withdrawn.Withdrawals may reflect toxicity, lack of efficacy, or economic reasons.
- Current impact in drug discovery: 49 publications in 2011 grew to 333 in 2020, reflecting the rapid expansion of AI4DD research.
3. AI applications in various stages of drug discovery
AI is being applied across disease characterization, target identification, molecular property prediction, generative chemistry, and safety assessment, but its reliability depends on data quality, novelty checks, and human oversight.
- 3.1. Diseases and therapy selection: AI supports disease diagnosis and classification, including image-based analysis in pathology, radiology, and dermatology.A CNN achieved near-physician accuracy for atopic dermatitis and distinguished it from several other skin conditions.
- 3.2. Target identification and validation: Knowledge graphs integrate heterogeneous biological entities and relationships into machine-learning-ready representations for target identification.They encode nodes such as genes, phenotypes, and compounds together with typed relationships and meta-paths.
- 3.2. Target identification and validation: XGBoost models using knowledge-graph data identified candidate genes for Alzheimer’s disease and autophagy, including previously unrecognized “ATG dark genes.”Among 251 predicted ATG-associated genes, 193 had no apparent connection to ATG, and literature review validated 7 of the top 20 and 2 of the bottom predictions.
- 3.4. Drug Safety: The road to clinical trials: AI-based drug-discovery models face limitations from arbitrary negative labels, data leakage, insufficient ground truth, regulatory transparency requirements, and inadequate compound novelty.Regulatory acceptance requires transparency about dataset origins, quality, and algorithms.
- 3.3. Hit generation and lead optimization: QSAR and related machine-learning methods predict bioactivity and ADMET properties from molecular structure, but rigorous analyses found GCNNs did not outperform classical QSAR fingerprints.QSAR spans target-based activity and properties such as solubility and permeability.
- 3.3. Hit generation and lead optimization: Generative chemistry methods such as GENTRL can produce candidate kinase inhibitors, while generated compounds require active filtering and human supervision.Filters remove reactive, chemically infeasible, or out-of-domain compounds; generated molecules may also be criticized for lacking novelty.
4. Brief overview of AI-driven drug discovery successes
AI-driven drug discovery has produced multiple compounds entering clinical development, spanning psychiatric, oncology, and fibrotic disease applications. Reported examples combine computational design with pharmacological selectivity and optimization.
- AI-designed clinical candidates: DSP-1181, a potent and long-acting 5HT1A receptor agonist, entered phase I studies for obsessive-compulsive disorder in January 2020.DSP-0039, a dual 5-HT1A agonist and 5-HT2A antagonist without dopamine D2 activity, also entered phase I studies for Alzheimer’s disease psychosis.
- Patent examples: ExScientia patents report compounds with shared pharmacological profiles but also compounds lacking chemical similarity to known drugs.Figure 3 contrasts gepirone and Example 1 with Example 109, illustrating chemical novelty alongside pharmacological comparison.
- Precision oncology: A single-cell functional precision medicine platform tested 139 drugs on samples from 143 patients with hematologic malignancies and guided treatment for 56 patients.Among treated patients, 30 (54%) experienced over 1.3-fold increased progression-free survival, while 12 responders had responses lasting three times longer than expected.
- AI-designed clinical candidates: ISM018-055 may have been the first AI-designed compound to enter phase II trials, targeting TNIK for idiopathic pulmonary fibrosis.The compound was designed using the PandaOmics and Chemistry42 platforms.
- AI-designed clinical candidates: RLY-1971 is an orally bioavailable allosteric PTPN11 inhibitor that blocks wild-type PTPN11 with IC50 <1nM and the E76K mutant with IC50 <250nM.It entered phase I trials for RTK/RAS-driven solid tumors and was licensed by Genentech in December 2020.
- AI-designed clinical candidates: RLY-4008 is a highly selective, potent, irreversible FGFR2 inhibitor effective in cholangiocarcinoma and progressed into phase I clinical studies.Its lack of potency against FGFR1 and FGFR4 may eliminate unwanted side-effects, although its exact bioactivity was not disclosed.
5. Challenges and limitations of AI in drug discovery
AI in drug discovery remains constrained by limited validation, unreliable or heterogeneous evidence, and the scale of chemical space. Effective use also depends on project-specific model selection and continuing human judgment.
- Scope of current AI: No regulatory-approved medicines can yet be attributed to a comprehensive AI system; reported successes mainly support human decision-making.The review distinguishes current computational assistance from end-to-end AI drug discovery.
- Chemical-space coverage: Virtual screening may involve 30 billion compounds and nearly 500 billion conformers, creating logistical and practical challenges beyond routine review scope.The review therefore emphasizes applicability-domain and external-predictivity validation.
- Chemical-space coverage: Models trained on tens of thousands of compounds or fewer may not adequately represent the chemical space of 30 billion compounds.This concern is especially important during lead optimization, where relevant chemical scaffolds must be accurately represented in the ML feature space.
- Data reliability: Drug-discovery evidence is affected by low reproducibility, including reported reproducibility rates of 33% and 11% for high-impact publications from Bayer and Amgen, respectively.Fabricated publications, inaccurate protocols, and weaker-than-published effects further increase concerns about the reliability of training data.
- Human and project-specific judgment: Selecting and ordering ML models depends on each project’s requirements, which may prioritize selectivity, tissue delivery, toxicity, permeability, or scaffold similarity.Different objectives can require filters, lead hopping, or sequential ML-model deployment.
- Human and project-specific judgment: Medicinal chemists continue to rely on judgment and may veto computationally acceptable compounds, making expertise, bias, and time constraints important influences on early discovery.The review presents the rule of five as an earlier example of integrating informatics into chemical-space narrowing.
6. Conclusions and future outlook
AI has produced drug-discovery successes, including compounds entering phase I trials, but it has not yet delivered comprehensive autonomous discovery. Its broader value depends on validated ground truth, careful model use, and human expertise.
- Current progress and remaining gap: AI-designed compounds have entered phase I clinical trials, but AI has not yet delivered a comprehensive system for fully automated drug discovery.These successes rely on computational platforms that support human decision-making.
- Human intervention and model use: Effective AI4DD use requires proper training and substantial human expertise for navigating complexity and supporting informed drug-discovery decisions.The review presents training as crucial for using AI4DD models effectively.
- Data quality and reproducibility: Ground truth across drug-discovery data types and relationships is necessary for AI systems to reveal their full strength.The review links this requirement to reproducibility concerns and the risk of processing questionable results.
- Human intervention and model use: Misguided use of AI4DD tools, such as assuming massive data alone produces actionable results, is identified as a potential danger.The authors instead describe AI steering much of the process, with human intervention at very late pipeline stages, as a better scenario.
7. Epilogue
A GPT-4 experiment illustrates AI’s ability to generate and modify chemically valid drug-like structures, while the review concludes that AI4DD has not yet reached its destination. The example remains a demonstration of potential rather than comprehensive autonomous discovery.
- Illustrative AI experiment: GPT-4 generated a valid chemical structure and modified Dasatinib into desmethyl-imatinib, a synthetically feasible kinase-inhibitor metabolite.The proposed molecule was available in the ZINC database and retained kinase-inhibitor relevance.
- Epilogue: The review concludes that AI4DD is progressing but is not yet fully mature.The epilogue frames the field as still advancing toward its goal.