AlphaFold at CASP13: First-Entry Win and Protein Folding

Twenty five out of forty three, on the first try

CASP is a biannual blind test: organizers send teams the amino acid sequences of proteins whose structures have already been solved by laboratory methods but not published, and teams submit predicted 3D shapes. Ninety eight groups entered CASP13. In the free modeling category, the hardest one, AlphaFold predicted the most accurate structure for 25 of 43 proteins. The next best team managed 3 of 43. [2]

Demis Hassabis, DeepMind's co-founder and CEO, called it the lab's first major investment of people and resources into a real world scientific problem, distinct from the games it had used to test its algorithms. Google had bought DeepMind four years earlier as a small research group with no shipped product. In between, DeepMind had beaten a champion Go player and then built a version of the same program that learned with no human game data at all. Hassabis's framing, in his own words, was that none of that had ever been about the games themselves. [2][1]

Atlas interpretation: The margin over second place is what made structural biologists sit up rather than the raw win. Mohammed AlQuraishi, a computational biologist who had entered his own method in the same competition, described the jump as roughly two CASPs' worth of progress compressed into one, measured against the field's own recent rate of improvement. A decade of near stagnation had only just started breaking in the two prior competitions, so a first time entrant beating that improved pace by a wide margin was the part worth explaining, not the fact that it won. [3]

Predict the distances, then fold to fit them

AlphaFold trained a neural network on known protein structures to predict, for a new sequence, the probability distribution of distances between every pair of amino acids and the angles of the chemical bonds connecting them. A second step turned those probabilistic distance predictions into a smooth energy function and then minimized it with gradient descent to arrive at a single 3D structure, working on the whole chain at once rather than piecing together fragments. [1][3]

Atlas interpretation: AlQuraishi's read on the method was that the optimization step, ordinary gradient descent over a learned energy surface, needed no elaborate folding pipeline once the distance predictions were good enough, and that this was closer to an engineering achievement built on existing co-evolutionary techniques than a wholly new idea. The Guardian reported the more visible sign of that engineering: DeepMind's first structure predictions took a fortnight to compute and later ones took a couple of hours. [3][2]

A first step, by the winner's own account

Hassabis told the Guardian, "We've not solved the protein folding problem, this is just a first step." Liam McGuffin, who led the top scoring UK academic group, said DeepMind had pushed the bar higher and that his own team, with far fewer resources, could still be competitive; he put full field wide progress on protein folding sometime in the 2020s. [2]

Atlas interpretation: AlQuraishi's caveat sharpens that modesty. The scores driving the headlines were GDT_TS, which measures whether the overall shape is right, and on the stricter GDT_HA measure of high resolution accuracy, useful for the drug design work everyone was already citing, the improvement looked far smaller. He also noted that a global score can hide local errors in exactly the regions, like an active site, that a biologist would care about most, and that a repeat of this scale of jump was not guaranteed since a lab of DeepMind's size and resources joins a competition like this only once. [3]