D-I-TASSER (Deep-learning based Iterative Threading ASSEmbly Refinement) is a new method extended from I-TASSER for high-accuracy protein structure and function predictions. Starting from a query sequence, D-I-TASSER first generates inter-residue contact and distance maps and hydrogen-bond (HB) networks using multiple deep neural-network predictors, including AttentionPotential (self-attention network built on MSA transformer) and DeepPotential. Meanwhile, it identifies structural templates by the meta-threading LOMETS3 approach, which includes the models built from the state-of-the-art AlphaFold2 program. The full-length atomic models are finally assembled by iterative fragment assembly Monte Carlo simultions under the guidance of I-TASSER force field and deep-learning contact/distance/HB restraints, where biological functions of the query protein are derived from the structure models by COFACTOR. The D-I-TASSER pipeline (as 'UM-TBM' and 'Zheng') was ranked as the No. 1 server/predictor in all categories of protein structure prediction in the most recent CASP15 experiment, including Multi-domain Targets, Single-domain Targets, and Multi-chain Targets. Please report problems and questions at our Discussion Board.
yangzhanglabumich.edu | (734) 647-1549 | 100 Washtenaw Avenue, Ann Arbor, MI 48109-2218