Accurate and efficient protein sequence design through learning concise local environment of residues

Abstract

Computational protein sequence design has been widely applied in rational protein engineering and increasing the design accuracy and efficiency is highly desired. Here we present an accurate and efficient design approach called ProDESIGNLE, which uses a concise but informative representation of residue’s local environment and trains a transformer to predict a residue’s amino acid type from its local environment. Using the trained transformer, ProDESIGN-LE iteratively selects an appropriate residue at a random position of the target structure, eventually acquiring a designed sequence with all residues fitting well with their own local environments.We assessed ProDESIGN-LE in silico by designing sequences for 68 naturally-occurring proteins used by the 14th critical assessment of protein structure prediction and 129 hallucinated structures. ProDESIGN-LE accomplished sequence design within 20 seconds per protein on average, and the predicted structures of these designed sequences perfectly resemble the target structures with state-of-the-art average TM-score (0.81 and 0.83 on these two datasets, respectively). We further experimentally validated ProDESIGN-LE by designing five sequences for an enzyme, chloramphenicol O-acetyltransferase type III (CAT III) and recombinantly expressing the designed proteins in E. coli. Of these designed proteins, three exhibited excellent solubility and one protein yielded monomeric species with circular dichroism spectra consistent with the natural CAT III protein. We anticipate that ProDESIGN-LE, with the improved accuracy and efficiency, would greatly facilitate rational protein engineering.

ThermoLink: Bridging Disulfide Bond and Enzyme Thermostability through Database Construction and Machine Learning Prediction

“…5A. Alanine has the highest frequency in both categories, and this is consistent with the BLOSUM62 matrix, indicating A has a high substitution probability with C. Interestingly, mutations that are more likely to be associated with improved enzyme thermostability involve conservative replacements, such as S and T, which may have less influence on local polarity and volume [73]. Typically, considerable proteins possess unbounded terminus and the stabilization of these unbounded short regions can greatly improve the overall thermostability of the enzyme.…”

All-atom protein sequence design based on geometric deep learning

“…Additionally, improvements were observed in the metrics of exposed hydrophobics. We further compared GeoSeqBuilder with representative deep learning sequence design approaches, including graph based ProteinMPNN [28], 3D-CNN based method [22], and transformer-based ABACUS-R [27] and ProDESIGN-LE [50]. The average pLDDT and TM-score (78.…”

From sequence to function through structure: Deep learning for protein design

“…MIF [70] adapted Ingraham’s [35] architecture to a bidirectional denoising model. ProDESIGN-LE [71] inputs structural local environments to three encoder layers to output a distribution over the 20 residue types. Graph Vector Perceptrons ( GVP-GNN ) [72] which replaced multilayer perceptrons (MLPs) in GNNs improved performance in protein design and model quality assessment.…”