Artificial Intelligence in Drug Discovery: Integrating Algorithms, Data, and Innovation for Accelerated Therapeutics

Drug discovery remains one of the most resource-intensive and time-consuming scientific endeavors. Developing a new therapeutic entity typically requires 10–15 years and investment exceeding US $2 billion, with fewer than 10% of candidates achieving regulatory approval (1,5). Traditional high-throughput screening and trial-and-error experimentation often result in poor hit rates and high attrition.

Artificial intelligence—particularly its subset, machine learning—has emerged as a disruptive force capable of addressing these inefficiencies by predicting bioactivity, toxicity, and pharmacokinetics with unprecedented precision (1–5,11). Its ability to simulate human reasoning, coupled with computational scalability, has enabled integration across the entire pharmaceutical pipeline—from target identification to drug repurposing (5,16,21).

Recent years have seen convergence of algorithmic innovation, big data availability, and cloud-based computation. Techniques such as deep neural networks (DNNs), graph neural networks (GNNs), transformers, and reinforcement learning (5,10,11,14) have become widely implemented. The success of AlphaFold (7), DeepTox and Deep Learning toxicity models (16,18), and ChemBERTa-type foundation models (10,24) exemplifies AI’s maturing role.

This review synthesizes key publications from 2019–2025 to provide a critical analysis of AI-enabled innovations and remaining challenges.

METHODOLOGICAL LANDSCAPE OF AI IN DRUG DISCOVERY:

Machine Learning and Deep Learning Foundations

Machine learning enables computers to learn predictive patterns from labeled or unlabeled data (5,11). Supervised learning predicts molecular activity or toxicity (11,16), while unsupervised learning uncovers intrinsic chemical clusters and patient subgroups (24).

Deep learning extracts hierarchical molecular and biological features using CNNs, RNNs, and transformers applied to images, graphs, and SMILES sequences (10,11,16).

Graph Neural Networks and Representation Learning

GNNs have become central to cheminformatics because molecules inherently form graph-structured data (10,18). Models such as GraphConv, SchNet, and MPNN capture structural properties for QSAR prediction and molecular property inference (10,16).

Generative and Reinforcement Learning Approaches

Generative models explore the vast chemical space—estimated at 10³?–10?? small molecules (10,14).

• VAEs and GANs generate novel bioactive structures (10).
• Reinforcement learning refines compounds by rewarding novelty and drug-likeness (14).

The GENTRL model produced potent DDR1 inhibitors in radically reduced timeframes (6,14).

Explainable and Ethical AI

Explainable AI (XAI) methods such as SHAP and attention visualization enable interpretation of deep-learning decisions (17,22,23). Ethical AI stresses fairness, accountability, and data privacy, particularly for clinical and omics-derived datasets (21,25).

APPLICATIONS ACROSS THE DRUG DISCOVERY PIPELINE:

Target Identification and Validation

AI models analyze multi-omics profiles to identify disease-relevant targets (24). Ensemble learning and network-based deep models help prioritize targets with high druggability (5,16). Bayesian and random forest classifiers support phenotypic correlation-based target selection (11,16).

Virtual Screening and Hit Discovery

Machine learning greatly enhances virtual screening accuracy (1,5,10). RF-Score and DeepDock outperform classical docking via learning-based affinity prediction (15,16). CNNs and GCNs extract spatial protein–ligand interaction patterns, dramatically reducing screening time (15,16).

Lead Optimization and Property Prediction

QSAR models and neural networks predict ADMET parameters (11,18). DeepTox and ADMET-AI apply DL-driven insights to early toxicity prediction (16,18). RL-guided optimization improves potency while maintaining drug-likeness (14).

Protein Structure Prediction

AlphaFold represents a transformative breakthrough, achieving near-experimental structural accuracy (7). AI-driven modeling extends to protein–ligand docking, binding-pocket identification, and mutant stability prediction (7,19).

Drug Repurposing

Models such as BioBERT and DrugNet extract drug–disease relationships from biomedical literature (22,23). AI-enabled repurposing accelerated antiviral candidate identification during COVID-19 (8,9,21).

Clinical Trial Optimization and Personalized Medicine

ML enhances clinical trials through risk prediction and patient-response modeling (24). EHR-driven early-warning systems have demonstrated predictive power for conditions such as sepsis (22,24).

ADVANCED TRENDS AND INTEGRATIVE FRAMEWORKS:

Large Language Models and Foundation Architectures

LLMs such as ChemBERTa, Galactica, and MolFormer integrate multimodal biomedical data (10,24). These models support property prediction, molecular captioning, and hypothesis generation.

Multi-Omics and Systems Biology Integration

ML-driven multi-omics fusion using autoencoders and graph networks correlates gene expression with drug responses (24). This underlies precision medicine and individualized drug design.

Emerging Therapeutic Modalities

AI assists in designing PROTACs, molecular glues, and RNA-targeting agents (16,19). DL models improve binding-kinetics prediction and enable targeting of undruggable proteins such as KRAS G12C (10,19).

Explainability and Regulatory Readiness

Regulatory bodies increasingly require interpretable models (17,25). XAI provides visualizations for toxicity substructures and efficacy determinants, improving medicinal-chemist confidence (17,22).

CHALLENGES AND LIMITATIONS:

Key limitations include:

1. Data Quality & Availability—restricted and imbalanced datasets (11,16).
2. Interpretability—deep networks remain black boxes (17,23).
3. Generalization—poor transfer to novel scaffolds (20,21).
4. Computational Cost—large models require significant hardware (19).
5. Ethical & Legal Issues—privacy and bias risks (21,25).

FUTURE PERSPECTIVE

AI will integrate with quantum computing, automated labs, digital twins, and federated learning frameworks (1,10,25). Explainability and hybrid human-AI workflows will guide trustworthy innovation (17,22,25). Democratized datasets will reduce global health disparities (1,3,24).

CONCLUSION

AI has become central to modern pharmaceutical R&D (1–5,13). Its integration across discovery pipelines accelerates development, expands molecular diversity, and enables targeting previously intractable proteins. However, interpretability, data governance, and ethical oversight remain essential (17,21,25). Responsible, transparent, and collaborative AI frameworks will drive the next generation of precision therapeutics.

REFERENCE

1. Ferreira FJN, Carneiro AS. AI-Driven Drug Discovery: A Comprehensive Review. ACS Omega. 2025; 10: 23889–23903.
2. Deng J, Yang Z, Ojima I et al. Artificial Intelligence in Drug Discovery: Applications and Techniques. arXiv:2106.05386 (2021).
3. Ugurlu SY, McDonald D, He S. Machine Learning Applications in Drug Discovery and Development. French-Ukrainian J Chem. 2025; 13(1): 1–66.
4. Rehman AU, Li M, Wu B et al. Role of Artificial Intelligence in Revolutionizing Drug Discovery. Fundamental Research. 2025; 5: 1273–1287.
5. Vamathevan J, Clark D, Chen E et al. Applications of Machine Learning in Drug Discovery and Development. Nat Rev Drug Discov. 2019; 18: 463–477.
6. Zhavoronkov A et al. Deep Learning Enables Rapid Identification of DDR1 Kinase Inhibitors. Nat Biotechnol. 2019; 37: 1038–1040.
7. Jumper J et al. Highly Accurate Protein Structure Prediction with AlphaFold. Nature. 2021; 596: 583–589.
8. Stokes JM et al. A Deep Learning Approach to Antibiotic Discovery. Cell. 2020; 180: 688–702.
9. Gao K et al. Generative Design of Novel Molecules for COVID-19 Using Graph Neural Networks. ChemRxiv 2020.
10. Walters WP, Barzilay R. Applications of Deep Learning in Molecule Generation and Property Prediction. Acc Chem Res. 2021; 54: 263–270.
11. Chen H, Engkvist O, Wang Y et al. The Rise of Deep Learning in Drug Discovery. Drug Discov Today. 2018; 23: 1241–1250.
12. Segler MHS et al. Planning Chemical Syntheses with Deep Neural Networks and Symbolic AI. Nature. 2018; 555: 604–610.
13. Walters WP, Murcko MA. Assessing the Impact of Generative AI on Medicinal Chemistry. Nat Biotechnol. 2023; 41: 141–153.
14. Popova M et al. Deep Reinforcement Learning for De Novo Drug Design. Sci Adv. 2018; 4(7): eaap7885.
15. Ragoza M et al. Protein–Ligand Scoring with Convolutional Neural Networks. J Chem Inf Model. 2017; 57: 942–957.
16. Rifaioglu AS et al. Recent Applications of Deep Learning and Machine Intelligence in Drug Discovery. Brief Bioinform. 2019; 20(5): 1878–1912.
17. Gao M, Pickett SD, Sumbalova L et al. Interpretable Machine Learning in Medicinal Chemistry. J Med Chem. 2022; 65: 10015–10029.
18. Chen L, Lu J, Zhang N et al. Predicting ADMET Properties with Graph Convolutional Networks. J Chem Inf Model. 2021; 61: 1563–1575.
19. Xu Y, Lin K, Wang S et al. Deep Learning for Free Energy Calculations in Drug Discovery. J Chem Theory Comput. 2021; 17: 6786–6798.
20. Altae-Tran H et al. Low Data Drug Discovery with One-Shot Learning. ACS Cent Sci. 2017; 3: 283–293.
21. Zhavoronkov A, Ivanenkov YA. Deep Learning for Drug Discovery: Challenges and Opportunities. Drug Discov Today. 2020; 25(12): 2221–2230.
22. Paul D et al. Explainable AI for Drug Repurposing and Mechanistic Discovery. Front Pharmacol. 2022; 13: 903849.
23. Jiménez-Luna J et al. Drug Discovery with Explainable Artificial Intelligence. Nat Mach Intell. 2020; 2: 573–584.
24. Zhang Q et al. Multi-Omics Integration for Precision Medicine in Drug Discovery. Trends Pharmacol Sci. 2023; 44: 275–289.
25. Vendrell JA et al. Ethical Considerations in AI-Driven Drug Development. Nat Med. 2024; 30: 332–340.