In a comprehensive discussion on the Latent Space podcast, prominent figures in computational biology offered a candid assessment of what artificial intelligence has truly achieved in protein research. Pushmeet Kohli, Head of AI for Science at Google DeepMind, and Sal Candido of the CZ Biohub examined the real capabilities of structural biology models like AlphaFold. Despite the global acclaim surrounding AlphaFold, both researchers stressed that the fundamental biophysical problem of protein folding remains unsolved. The widespread public perception that the folding problem has been settled fundamentally misunderstands the reality of living biology.
According to Kohli and Candido, the primary limitation lies in the nature of the training data and prediction targets. AlphaFold essentially reconstructs the static conformations cataloged over decades in the Protein Data Bank. As a result, the model predicts rigid three-dimensional snapshots of stable end states rather than simulating the actual folding process as a time-dependent physical sequence. In nature, amino acid chains navigate complex energy landscapes, moving through transient conformations that static coordinate models cannot capture.
Real biological macromolecules inside cells are rarely rigid sculptures. Natural proteins are flexible, highly context-dependent, and constantly shifting their shapes in response to environmental cues, pH levels, and neighboring molecules. Moreover, a substantial fraction of proteins are intrinsically disordered, meaning they lack a single stable structure when isolated. These unstructured molecules only adopt specific conformations when interacting with particular binding partners, rendering purely static predictions inadequate for understanding their functional roles.
These constraints carry significant implications for modern drug discovery. Real breakthroughs in designing therapeutic compounds require a detailed understanding of how target molecules behave dynamically and how binding events trigger conformational changes. Knowing a single frozen crystal structure is rarely sufficient to predict how a candidate drug will affect cellular pathways. Because current models miss these temporal transitions, researchers trying to engineer precise pharmacological interventions encounter severe functional bottlenecks.
To overcome these barriers, Kohli and Candido argue for an ambitious shift toward modeling entire cellular systems, often referred to as a Virtual Cell. Such an initiative must move beyond isolated individual proteins to simulate molecular dynamics, cellular functions, and multi-protein networks concurrently. Building a functioning virtual cell requires generating new types of biological data that capture motion and context, rather than relying exclusively on historically accumulated crystallographic databases.
The researchers linked this strategic pivot directly to the Bitter Lesson of artificial intelligence, which posits that general methods leveraging scalable computation consistently outperform human-designed domain heuristics. In computational biology, future breakthroughs will likely stem from providing massive compute capacity to broad datasets of cellular dynamics rather than hand-tuning structural rules. Only by scaling computation across dynamic biological environments can AI deliver on its full potential for medicine.