Interview
AI in Pharmaceutical R&D with Kim Branson
- Kim Branson projects that AI-driven "virtual humans" fully replacing clinical testing in drug discovery is approximately 50 years away, while predicting that within five years, every drug launched by GSK will be supported by software predicting patient response and dosage.
- The outlook anticipates that within five years, the industry will achieve a significantly deeper understanding of immunotherapy (IO) response rates, currently at 20%, and will move from generating many experiments to conducting fewer but more informative ones using large-scale perturbational data as inference tables.
- A shift toward "systems medicine" is expected, characterized by the integration of mechanistic modeling of biology and structured priors regarding organ systems into machine learning, moving beyond reliance on gene expression alone.
- The availability of "data with outcomes" is identified as the current rate-limiting step for machine learning advancement, necessitating the establishment of large-scale public-private consortia and long-term observational cohorts that follow people over time to capture immune system dynamics.
- Future drug development plans include the creation of long-term biobanks containing physical samples alongside genetic sequencing, leveraging "Salkers law" where analysis becomes cheaper and better to enable retrospective analysis and comprehensive data collection.
- GSK expects exponential data growth from the emerging age of single-cell and perturbational genomics, which will facilitate the deciphering of disease heterogeneity and the identification of specific patient subsets capable of functional cure.
- Computational pathology and active learning systems are predicted to drive precision, enabling the continuous modulation of gene expression, precise identification of responsive cell types, and the use of continuous traits for imaging data rather than binary or manual scoring.
- Algorithmic requirements will evolve to include well-characterized robustness, reliability criteria, and confidence measures rather than point estimates, with a continued appreciation for simple algorithms like Random Forests when paired with sufficient unique data.
- The industry faces the challenge of organizational alignment and communication complexity, described as "turning the ocean," with a strategic shift expected from "dabbling" to establishing AI as a core strategy requiring flexible integration options and user-friendly tools.
- Risks and operational constraints include the need to minimize patient burden while measuring everything in clinical trials, the challenge of model drift, and the necessity of ensuring AI tools are robust enough for clinical adoption before widespread integration.
- Machine learning capabilities are forecast to expand to predicting directionality for genetic variants, identifying subsets of people for functional cures through viral surface antigen reduction, and automating the search for scientific answers rather than document retrieval.
- Future AI adoption in biopharma will rely heavily on unique data generated by companies rather than public data or personnel alone, with the ultimate goal of inferring experimental outcomes to reduce the physical conduct of experiments.