Pushmeet Kohli

Pushmeet Kohli is a Google DeepMind researcher whose work connects machine learning with problems in biology and other sciences. A co-author of the AlphaFold research program, he emphasizes both the practical value of predicting a protein structure and the biological questions a single prediction leaves open.[1][2][3]

In the 2021 AlphaFold account, Kohli and colleagues describe a model that takes an amino-acid sequence and related sequence information, builds a representation of relationships between residues, and predicts three-dimensional coordinates. Kohli connects that design to biological knowledge: residues influence one another, so a model that represents their relationships can use data more efficiently than one that must rediscover everything unaided.[3][1]

Kohli stresses the distinction between that result and a complete account of protein biology. AlphaFold was trained to predict structures obtained and deposited by researchers; proteins can be disordered or adopt different shapes in different contexts. A useful predicted structure is therefore not a full distribution of molecular states, a description of how the protein moves, or a direct measurement of what it does in a cell. Those remaining questions matter for understanding function and for testing new protein designs.[1]

One possible source of richer information is cryo-electron microscopy. Kohli describes trying to work directly from cryo-EM micrographs—the underlying images used in structural analysis—in the hope of learning about distributions and dynamics. He presents that work as unfinished, not as a solved route to molecular motion.[1]

Kohli argues that useful predictions require a reliable sense of uncertainty. A confidently wrong prediction can send a laboratory down an expensive path, so AlphaFold’s confidence estimates help users judge which regions deserve trust. He distinguishes this behavioral understanding—knowing where the model tends to work or fail—from explaining every internal step that produced an answer. Neither confidence nor a plausible structure replaces experimental validation.[1][3]

His wider approach starts with the scientific problem rather than a preferred technique. He contrasts AlphaFold, which could build on an existing investment in protein structure data, with attempts at a virtual cell where the necessary data were not yet sufficient. The choice may be to improve the model, gather new measurements, or combine both with scientific expertise.[1]

Kohli began in security research and moved through computer vision and machine learning into AI for science. Google identifies him as a DeepMind vice president and Google Cloud chief scientist. In drug discovery he separates AI’s existing use in parts of the process from claims of a large acceleration in clinical timelines: the latter depends on which stage is being accelerated and on better biological models.[1][2]