AI in drug discovery highlights data gaps and lab bottlenecks despite early promise

AI is being adopted to compress timelines and reduce risk in early-stage drug discovery, particularly in hit identification, where models design and screen molecular candidates faster than ever. According to Paul Belcher, director of protein research strategy at Cytiva, the main cost in drug discovery remains the clinical phase, making earlier-stage risk reduction a key priority. AI’s ability to predict interactions and eliminate low-quality candidates before physical testing is seen as a major efficiency gain, but it does not yet reliably predict compound kinetics or developability, requiring lab validation of every AI-generated candidate.

The shift from empirical to predictive design has increased pressure on lab teams, which now face validating, characterizing, and purifying a growing volume of more diverse, AI-generated compounds. Traditional screening workflows, optimized for binary yes-or-no responses at scale, struggle to profile complex candidates in detail, exposing a mismatch between computational prediction and physical validation capabilities.

Industry data quality issues are compounding these challenges. Many AI models trained on public datasets are hitting a 'data wall,' where limited and repetitive data leads to diminishing returns and model bias. Public datasets and scientific publications disproportionately favor positive results, leaving models without comprehensive negative data—failed experiments or non-binding compounds—that would improve reliability. Belcher notes that negative data is often buried in lab notebooks and rarely shared, reinforcing publication bias and limiting model robustness.

Data integrity concerns are rising alongside the use of AI in research. Fabrication and manipulation of scientific images, such as Western blots, have been documented, and generative AI has made such misconduct easier. Belcher cites research identifying nearly 4% of biomedical papers containing duplicated or manipulated images, underscoring the need for verification tools to ensure data authenticity before it is used to train models.

Solutions are emerging to address these gaps. Vendors like Cytiva are developing tools such as Image Integrity Checker, which uses secure hash algorithms to detect tampered scientific images. Publishing houses are reportedly adopting such tools to verify data integrity in submitted work. These steps aim to restore trust in the data pipeline that feeds AI models in drug discovery.

The long-term vision for AI in drug discovery includes fully autonomous labs that operate with minimal human intervention, cycling continuously between prediction, testing, and optimization. Belcher describes these 'dark labs' as dependent on consistent, interoperable data and infrastructure. However, most labs today operate with standalone instruments and closed ecosystems, impeding the flow of data needed to close the loop between computational prediction and experimental validation. Achieving FAIR (findable, accessible, interoperable, and reusable) data at scale remains a prerequisite for this autonomous future.