with @vijaypande @smc90 We cover the news around Google DeepMind's AlphaFold system outperforming all others in the 14th biennial CASP challenge, which tracks progress, key metrics, and state-of-the-art on predictive techniques for 3-D protein folding. Structure is function in biology, and this is one of the grand challenges in biology. So is it really solved? Will it really revolutionize drug discovery? What's hype/ what's real when it comes to the buzz here; what are other applications; and what are the implications for open science, scientists, big companies, startups?
In this episode of our show 16 Minutes -- where we talk about the headlines, and where we are on the long arc of tech trends -- we cover the news around Google DeepMind's AlphaFold system for predicting the 3-D structure of proteins outperforming 100 teams across 20 countries in the 14th Community Wide Assessment on the CASP (Critical Assessment of Structure Prediction) challenge. The challenge, which takes place every other year (over several months) tracks progress, key metrics, and state-of-the-art on predictive techniques for protein folding.
This isn’t just an academic challenge; it matters because proteins define and power ALL life functions, and as the saying goes, “structure is function”: Figuring out the shapes that proteins assemble into is important in helping determine their functions and therefore potential applications (drug discovery, among other things). However, the astronomical number of possible structures for proteins -- and difficulty of figuring out these out (whether experimentally or computationally) from their amino-acid sequences alone -- has made it one of the grand challenges in biology. Some of the older techniques are described as "kind of like making a finger puppet to cast a shadow, and then trying to figure out what your fingers were like from the shadow"...
So is this grand protein folding problem really solved? Will it really revolutionize drug discovery? What's hype/ what's real when it comes to the buzz here; what are other applications; and what are the implications for open science, molecular biologists, computer scientists; big companies, startups? General partner Vijay Pande -- formerly professor of chemistry and structural biology and computer science, among other things at Stanford -- also founded the Folding@home project (which pioneered using distributed computing to solve the protein folding problem) and chats with Sonal Chokshi about whether this is a breakthrough or not. What is it, and where are we, really... ImageNet moment? E-MC2? Internet 1.0? Woodstock?!
"‘It will change everything’: DeepMind’s AI makes gigantic leap in solving protein structures", Ewen Callaway, Nature, 30 November 2020
"‘The game has changed.’ AI triumphs at solving protein structures", Robert Service, Science magazine, 30 November 2020
"DeepMind’s protein-folding AI has solved a 50-year-old grand challenge of biology", Will Heaven, Technology Review, 30 November 2020
images/ source: median accuracy in free-modeling category over the past 14 years; two protein targets and AlphaFold predicted structures compared against experimental result, both from the free modeling category / DeepMind