A NOVEMBER TO REMEMBER

Dick Pountain /Idealog 317/ 04 Dec 2020 10:43

World-changing innovations are like London buses: you wait for ages and then three come along at once. The recent wait has been particularly irksome, as under pandemic lockdown virology and epidemiology felt like the only relevant sciences – apart from rocket science, to get us all off this pestilential planet and to Mars (joke). Then suddenly, blam, three arrive in the same month: I’m writing this at the beginning of December 2020.

Most important and welcome was the arrival of not one but three coronavirus vaccines, all produced in record-breaking time and apparently highly effective (though that will only become certain once they’ve been deployed widely for a while). While their speed of development, testing and approval is remarkable, that’s not the innovation I meant though. The innovation is that two of the three are totally synthetic. Most vaccines up till now have required the target virus to be cultured in animal cells, then deactivated or broken apart and bits injected into live animals to generate antibodies. The Pfizer and Moderna vaccines are for the first time purely chemical: strings of Messenger RNA assembled from scratch by those sophisticated gene-sequencing and nucleotide assembling machines that have been invented over recent decades by harnessing powerful computers to robotic chemical processing. Injected into your arm, this mRNA tells your own cells to start making one harmless spike-protein of the coronavirus, a sort of pretend infection that generates antibodies against the real one.

The second innovation is in the same area of molecular biology, namely a great breakthrough in solving the ‘protein-folding problem’. All the biological processes that animate living things are driven by enzymes, which are proteins - very long chains made up from 20 or so different amino acid units. These chains don’t just flap around like pieces of string but fold themselves into compact lumps whose exact 3D shape, their cavities and crevices, enable them to work by fitting the molecules they work upon in lock-and-key fashion. These lumps are held together by bonds between amino acids from different points in the chain, and to design artificial enzymes, or drugs that alter the action of natural enzymes by fitting their slots, then you must simulate the way any particular chain of amino acids will fold itself. Since every link in the chain can rotate freely this is a crushingly difficult computational task if the chain is thousands long.

Back in October 2002, PC-owning nerds were being encouraged to donate their spare CPU cycles to a world-wide distributed network for solving protein-folding problems run by Stanford University. 30,000 donors helped crack protein structures around 50 amino acids, but 100 amino acids would need another 270,000. On 30th November 2020 an article in Nature related how Google’s AI subsidiary DeepMind – famous for cracking the board game Go sufficiently to beat the human world-champion – had made a similar breakthrough in protein-folding simulation.

DeepMind’s program AlphaFold 2 had outperformed around 100 other teams in a biennial protein-structure prediction challenge called CASP (Critical Assessment of Structure Prediction). The performance of folding prediction programmes is measured against physical  experimental evidence of a protein’s structure from X-ray Diffraction Crystallography or the newer Cryo-Electron Microscopy, which pictures the individual atoms: AlphaFold 2 scored close to a 90% match in this challenge. An earlier version of AlphaFold worked entirely by deep-learning, that is by examining the amino acid sequences versus 3D structures of many proteins, which enabled matches of around 60-70% accuracy. For AlphaFold 2 the team added an extra level, not of deep-learning but of constraint-solving. Consider any pair of amino acids linked together and physical chemistry can tell you how they can rotate about that bond and what resistance will be encountered: similar data is available for the close approach of active groups on remote amino acids. Applying such constraints to a purely learned prediction can boost the accuracy to 90%. The implications for the study of human biochemistry, disease and the speeding of up future drug design are massive, potentially world-changing.

The third innovation unveiled in November relates tangentially to the other two: Apple released its M1 CPU chip, which has two perhaps world-changing virtues. Firstly it breaks Apple’s own dependency upon Intel and more importantly elevates the ARM architecture, which already owns the mobile market, to desktop PC status, thus threatening Intel’s x86 hold there. (Ironically enough our dimwitted government recently allowed ARM to be sold abroad). Secondly the M1 chipset contains not only eight general-purpose ARM cores and an 8-core graphics unit, but also a 16-core Neural Engine capable of performing up to 11 trillion deep-learning operations a second. Alongside those gene-sequencing and nucleotide assembling machines, M1-powered computers running AlphaFold-style software promise a new era of computer-aided biology, a sort of Lego with living cells.

[Dick Pountain would like an M1-powered Chromebook for Christmas (perhaps not this one)]