AI-Driven Drug Discovery and Translational Medicine: From Computational Models to Clinical Impact

Abstract

Background: Artificial intelligence (AI) and machine learning are increasingly being used to extract signals from chemical data and high-dimensional biomedical, and to support data-driven prioritisation throughout the process of taking a discovery to the clinic. Objective: The study aims to map the application of AI and machine learning methods across the drug discovery–translational pipeline, and to evaluate the impact on clinical practice and the barriers to implementation. Method: A scoping review of peer-reviewed studies (2015–2025) was conducted using Embase and PubMed. Eligible studies reported real-world AI/ML applications with measurable outputs and applied validation (in vivo, cellular, biochemical, clinical, or trial-operational). Data were synthesized thematically and charted using an eight-domain extraction framework. Results: A total of 19 studies were included in the analysis. The evidence was found to be primarily focused on hit discovery (DNA-encoded library learning and virtual screening), preclinical modeling, lead optimization, and de novo/generative design. A smaller subset of studies addressed clinical trial operations and peptide/biologic antimicrobial discovery. Most studies reported prospective experimental validation, often involving animal models. However, only a minority demonstrated direct clinical translation, such as an AI-discovered candidate that was evaluated in a randomized phase 2a trial. Meanwhile, trial-operations tools showed consistent workflow gains (e.g., reduced time and increased screening accuracy). Key limitations included limited interpretability, uncertainty reporting, weak external validation, and restricted data access and representativeness. Conclusion: The application of AI/ML can improve trial processes and accelerate early discovery, but achieving consistent clinical impact remains constrained by governance, evaluation rigour and data quality. Conducting prospective multi-site evaluations, creating benchmarked and uncertainty-aware models, and strengthening open data practices are essential for translation.