DLPacker: Deep learning for prediction of amino acid side chain conformations in proteins

Abstract

Prediction of side chain conformations of amino acids in proteins (also termed "packing") is an important and challenging part of protein structure prediction with many interesting applications in protein design. A variety of methods for packing have been developed but more accurate ones are still needed. Machine learning (ML) methods have recently become a powerful tool for solving various problems in diverse areas of science, including structural biology. In this study, we evaluate the potential of deep neural networks (DNNs) for prediction of amino acid side chain conformations.We formulate the problem as image-to-image transformation and train a U-net style DNN to solve the problem. We show that our method outperforms other physicsbased methods by a significant margin: reconstruction RMSDs for most amino acids are about 20% smaller compared to SCWRL4 and Rosetta Packer with RMSDs for bulky hydrophobic amino acids Phe, Tyr, and Trp being up to 50% smaller.

Cation−π Interactions and Their Role in Assembling Collagen Triple Helices

“…Each dimension for each experiment was Fourier-transformed and phase-corrected manually using the TopSpin 4.0 software. NOE correlation was rendered using DLPacker and then visualized in PyMol. , …”

“…NOE correlation was rendered using DLPacker and then visualized in PyMol. 41,42 MD Simulations. All-atom MD simulations were utilized to understand side-chain interaction effects on the stability of the OGXY′ and YGXO peptides, where X was substituted with Phe (F) and Trp (W), and Y was substituted with either Lys (K) or Arg (R).…”

Protein Function Analysis through Machine Learning

“…Eventually, neural networks were applied to recognize protein folds [ 202 ] or to self-improve a predicted folded structure [ 203 ]. Subsequently, predicting optimal sidechain packing was addressed in pytorch implementations [ 204 , 205 ]. Using specialized transformations, such as Voronoi tessellations, and CNN has been shown to yield state of the art fold modeling [ 206 ].…”

Exploiting Sequence-Dependent Rotamer Information in Global Optimization of Proteins

“…A rotamer entry usually specifies the amino acid, the dihedral angles, with an associated measure of variance, and the probability of occurrence . Many rotamer libraries have been constructed and have been used in applications such as crystallographic model building, − protein–ligand docking, − homology modeling, − and protein design. − Within these applications, it is also possible to use machine learning to predict the most probable rotamer for a given conformation. − Moreover, the native structure of many proteins can now be predicted at atomic accuracy by neural networks, but there remain numerous peptide classes with little experimental data and important cases where we require additional minima beyond the native conformation. One conformation is insufficient for sampling the thermodynamic properties of the folding funnel and for predicting competing conformations and their transition rates.…”

“… 17 − 23 Within these applications, it is also possible to use machine learning to predict the most probable rotamer for a given conformation. 24 − 26 Moreover, the native structure of many proteins can now be predicted at atomic accuracy by neural networks, 27 but there remain numerous peptide classes with little experimental data and important cases where we require additional minima beyond the native conformation. One conformation is insufficient for sampling the thermodynamic properties of the folding funnel and for predicting competing conformations and their transition rates.…”