AI for Bio has a Fuzzy API problem (ankitg.me)

The recent commentary on the challenges of applying artificial intelligence (AI) in drug discovery highlights a critical issue known as the "fuzzy API problem" in biology. Unlike clean software interfaces that allow for straightforward interactions between components, the biological processes involved in drug development are complex and laden with uncertainties. For instance, target discovery often results in probabilistic hypotheses rather than definitive outcomes, while drug design may yield candidate molecules that require extensive validation across various biological contexts before being deemed viable. The clinical trial outcomes are similarly influenced by numerous variables, leading to a lack of clear-cut relationships between steps in the process. This fuzziness significantly complicates the integration of machine learning (ML) in bio-related fields. While ML shows promise in all stages—target discovery, drug design, and clinical development—the inherent complexity of biological systems often leads to challenges that could undermine AI’s effectiveness. The commentary suggests that overcoming these difficulties requires not just better algorithms but also improved data quality, innovative experimental designs, and enhanced causal inference capabilities. Companies exploring "virtual human" models may hold the key to making meaningful advancements, yet the journey from theoretical models to actionable insights remains fraught with complications. As such, addressing the fuzzy API problem represents both a challenge and a unique opportunity for the AI/ML community in revolutionizing biomedicine.