Bristol Myers Squibb

BMS faced the challenge of inefficient drug discovery where molecules were synthesized in the lab before being validated for efficacy, safety, and other properties. This funnel-based screening approach wasted resources on redundant or incremental molecule candidates, while looming patent cliff losses on blockbusters like Revlimid, Pomalyst, and Sprycel increased pressure to de-risk and accelerate R&D.

BMS developed and implemented proprietary AI/ML tools under a 'predict-first' model that evaluates molecule candidates for efficacy, safety liabilities, and other properties computationally before any lab synthesis occurs. The approach shifts from funnel-based screening to a 'choose your own adventure' model where scientists use AI-guided predictions to strategically prioritize which molecules to make and how to test them, including applying the approach to CELMoD agent programs for sickle cell disease.

All BMS small molecule programs now use AI to evaluate properties prior to synthesis, up from just 5% in 2021. The approach is also being applied to nearly half of large molecule experiments. In a concrete example, the sickle cell disease CELMoD program had plateaued for months before computational tools were applied — within weeks the team crossed the plateau and successfully combined multiple required parameters into a single molecule candidate.

• Shifting from lab-first to predict-first can dramatically accelerate drug discovery timelines, as demonstrated by breaking a months-long plateau in weeks on the sickle cell CELMoD program.
• Adoption requires a cultural evolution: increasing the ratio of computational scientists to lab experimentalists and embedding AI into team composition and decision-making processes.
• The predict-first model doesn't eliminate uncertainty but strategically narrows the search space, enriching the quality and strategic value of molecules that proceed to the lab.