AI Can Predict a Protein. Turning That Prediction into a Therapy Still Requires Laboratory Work, Manufacturing, and Patient Capital: Dorothy Chou’s Argument
AI Can Predict a Protein. Turning That Prediction into a Therapy Still Requires Laboratory Work, Manufacturing, and Patient Capital: Dorothy Chou’s Argument
On July 16, Dorothy Chou, the former head of public engagement at Google DeepMind, explained why a computational breakthrough does not, by itself, accelerate biology. She argues that funding should cover the entire path from prediction through validation and manufacturing.
In a conversation with the Foresight Institute, Chou begins with AlphaFold. The system predicts a protein’s three-dimensional shape from its sequence. Knowing that shape helps researchers determine where a drug molecule or antibody might bind to the protein. The model has made structures available for more than 200 million proteins. A long chain of work still separates a protein structure displayed on a screen from a treatment that can be given to a person.
AlphaFold’s success depended on open data and a shared method for validating results. The Protein Data Bank is a global archive of experimentally validated protein structures. CASP is a blind test in which teams predict structures that have already been determined experimentally but have not yet been published. In their paper on AlphaFold2, the researchers reported accuracy close to experimental results for most proteins in CASP14. This validation framework made it possible to compare models by accuracy rather than by how persuasive their presentations appeared.
The expensive work begins after such a prediction. Researchers must test the candidate in a wet laboratory, where they physically work with cells and tissues, and then establish a manufacturing process. Physical validation and manufacturing require time, infrastructure, and capital. Chou identifies this gap as the main obstacle that remains after the computational stage.
According to Chou, a venture capital fund typically expects to exit an investment within 5–10 years, while public funding often divides biology into narrow disciplines and annual budget cycles. She proposes blended financing. Grants and philanthropy would fund the long and risky part of the process, while private capital would scale candidates that have passed validation. Advance purchase commitments for vaccines during COVID-19 provide an example. Governments reduced manufacturing risk, which allowed companies to move faster.
For the next major idea, Chou proposes assembling the necessary data, independent validation, laboratory capacity, manufacturing, and suitable financing in advance. This would allow AI to move successful predictions through validation and into practical use instead of leaving them on a screen.