Biomedical Technology · global
How Far Has AI Drug Development Come? A Review of 57 Studies Reveals the Gap Between Model Performance and Therapeutic Efficacy
A systematic review examining pharmaceutical research after the pandemic found that AI applications remain concentrated in molecular screening and early-stage design. Predictive scores have shown considerable progress, but whether AI can shorten development and improve outcomes for patients still lacks sufficient validation.
Selecting molecules worth testing from vast compound libraries is one of the most demanding starting points in drug development. Artificial intelligence is helping researchers narrow the search, but how far is accurate computational selection from drugs that can treat disease safely and effectively? A new systematic review of 57 studies finds that current evidence mainly supports AI’s value as an aid in early discovery, while improvements in clinical efficacy or patient health outcomes have yet to be demonstrated.
The study was published on October 30 in *Frontiers in Artificial Intelligence* and was conducted by Xin Yi Chang, Clarence Teo and Cheong Hian Goh. The team identified 593 records in PubMed and Scopus and selected 57 studies published from 2023 to 2025. The search included COVID-19-related keywords, so the findings reflect research within a specific search scope and cannot be taken as a complete picture of AI adoption across the pharmaceutical industry.
Among these papers, drug discovery and design accounted for 54.4% of research areas, while deep learning was the most common technology category, at 35.1%. Specific tasks included predicting molecular binding to proteins, screening compounds for potential activity and exploring new uses for existing drugs. Many studies focused on coronavirus-related targets, using models to rank candidate molecules before subsequent experimental testing.
More revealing than the number of applications is how researchers measured their results. The review recorded 152 mentions of evaluation metrics, of which 80.9% were measures of AI or machine learning model performance, such as classification accuracy and prediction error. The denominator for this percentage is the number of metric mentions, not the 57 studies; a single paper could use multiple metrics. Reports on the same review by Bioengineer.org and Scienmag also emphasized this basis for the calculation.
These scores can help compare models, but they are insufficient to answer what will happen when a candidate drug enters the human body. The authors noted that most included studies were rated as having a high risk of bias and lacked external validation, making their findings more suitable for generating hypotheses to be tested. Whether models can maintain their performance on data they did not encounter during development, and whether computational advantages can translate into shorter development times, lower costs or better therapeutic efficacy, all require external validation and prospective studies.
Background Context
The evidence chain for AI drug development begins with data and extends to experiments and humans. The review noted that some models were trained and validated using compound data from databases such as PubChem and BindingDB. Whether those data are reliable, and whether predictions can extend beyond the original data’s scope, are questions that must be asked when interpreting the results. The authors suggested that future work could further incorporate data on how drugs move through and act within the body, bringing model evaluation closer to actual medication use.
The review itself also has limitations: it searched only two databases, covered three years of literature and did not preregister its review protocol. Positive findings are also more likely to be published, potentially amplifying the impression of technological benefits. Its contribution is to place the results of AI drug development within a testable framework: molecular screening has produced usable leads, but evidence of how much these leads can benefit patients must still be gathered step by step along the drug development process.