Structure prediction has been transformed by deep learning, but large multi-domain proteins remain stubborn. Such proteins consist of several semi-independent sections whose relative arrangement is poorly constrained by sequence alone, and purely learned models tend to produce confident predictions for each domain and unreliable guesses about how they fit together. A tool announced by NUS, D-I-TASSER, is built specifically for this regime.
Its strategy is divide and conquer, reinforced by physics. Deep learning predicts the shape of each protein section, and physics-based simulation then assembles those sections into a complete structure subject to energetic plausibility. The combination addresses the characteristic failure mode of learned methods, which is that the training data contain comparatively few examples of full-length multi-domain architectures from which inter-domain geometry could be inferred.
The reported gains are concrete. D-I-TASSER predicted complex protein structures about 13 per cent more accurately than current leading methods, and the team generated reliable models for most of the roughly 20,000 proteins in the human body. The work was led from the Cancer Science Institute of Singapore with appointments in the NUS School of Computing and the Yong Loo Lin School of Medicine.
Planned extensions indicate where the method is headed: RNA structure prediction, protein and protein interactions including antibody and antigen complexes, and modelling how proteins fold inside cells rather than in isolation. The last of these is the most ambitious, since cellular folding involves chaperones and crowding effects that neither learned nor physics-based approaches currently capture well.