For years, much of the conversation around AI in drug discovery has focused on prediction: identifying better targets, designing stronger molecules and improving the accuracy of computational models.
The next phase may be fundamentally different.
The biggest opportunity for AI could come from connecting discovery activities that have traditionally operated in sequence — computational design, experimentation, biological assessment, optimisation and developability — into systems that continuously learn from one another.
That is the direction highlighted by Richard Bonneau, VP & Global Head of AI for Drug Discovery at Roche and Genentech, in a recent Oxford Global thought leadership interview.
AI is not new to drug discovery. Computational methods have supported molecular modelling, genomic analysis and target prioritisation for years. What is changing is the emergence of increasingly capable reasoning and orchestration systems that can combine multiple analyses, models and scientific workflows.
This creates the possibility of moving beyond isolated AI tools towards a much more connected discovery engine.
A key part of this is the concept of the lab-in-the-loop.
Experimental data can be fed directly back into computational models, allowing each round of laboratory work to improve the next set of predictions. Instead of a linear process, discovery becomes iterative: design, test, learn and redesign.
The opportunity becomes even greater when learning extends across multiple programmes. Insights generated in one project can potentially strengthen models used elsewhere, allowing large R&D organisations to extract more value from both current and historical experimental data.
But the most important change may be in decision-making.
AI systems are increasingly being developed not only to predict what might work, but to help determine what scientists should test next. Active learning approaches could identify the experiments most likely to generate useful information, reducing unnecessary cycles and making expensive laboratory work more targeted.
Combined with automation, this starts to create a very different model for discovery.
The challenge, however, is integration.
Drug discovery is not one prediction problem. It is a chain of interconnected scientific and operational decisions, each carrying its own uncertainty. A highly accurate model at one stage offers limited value if its output fails when transferred into the next.
The real competitive advantage may therefore come from building AI systems that understand the wider discovery process — incorporating biology, molecular properties, developability and downstream requirements from the outset.
As Bonneau puts it, one of the most exciting areas now is the integration of AI-driven design with automation.
That points towards a broader evolution in the role of AI.
The future may not belong to the company with the single best algorithm. It may belong to the organisations that can connect AI, experimental science, automation and institutional knowledge into a continuously improving discovery system.
