Why AlphaFold Didn't Solve Protein Folding — Pushmeet Kohli, Google DeepMind & Sal Candido, Biohub

The "bitter lesson"—the principle that scalable methods eventually outperform handcrafted ones—requires a critical re-evaluation when applied to data-driven bioscience. While scaling compute and data is essential, success depends on identifying the right problem and prioritizing data quality over mere quantity. Modeling efforts, such as AlphaFold, demonstrate that integrating scientific intuition with machine learning creates more data-efficient solutions than brute-force scaling alone. True progress in translational medicine necessitates a multidisciplinary approach where researchers move beyond rigid adherence to either modeling or data generation. Future breakthroughs in protein dynamics and function rely on transitioning from static structure prediction to capturing complex biological interactions, ultimately requiring a shift toward 10x-impact goals rather than incremental improvements. This evolution demands open collaboration and a focus on building trustworthy models that provide actionable insights for clinical applications.