An official website of the United States government
Abstract Machine learning has been increasingly used for protein engineering. However, because the general sequence contexts they capture are not specific to the protein being engineered, the accuracy of existing machine learning algorithms is rather limited. Here, we report ECNet (evolutionary context-integrated neural network), a deep-learning algorithm that exploits evolutionary contexts to predict functional fitness for protein engineering. This algorithm integrates local evolutionary context from homologous sequences that explicitly model residue-residue epistasis for the protein of interest with the global evolutionary context that encodes rich semantic and structural features from the enormous protein sequence universe. As such, it enables accurate mapping from sequence to function and provides generalization from low-order mutants to higher-order mutants. We show that ECNet predicts the sequence-function relationship more accurately as compared to existing machine learning algorithms by using ~50 deep mutational scanning and random mutagenesis datasets. Moreover, we used ECNet to guide the engineering of TEM-1 β-lactamase and identified variants with improved ampicillin resistance with high success rates.
more »
« less
Deep learning the structural determinants of protein biochemical properties by comparing structural ensembles with DiffNets
Abstract Understanding the structural determinants of a protein’s biochemical properties, such as activity and stability, is a major challenge in biology and medicine. Comparing computer simulations of protein variants with different biochemical properties is an increasingly powerful means to drive progress. However, success often hinges on dimensionality reduction algorithms for simplifying the complex ensemble of structures each variant adopts. Unfortunately, common algorithms rely on potentially misleading assumptions about what structural features are important, such as emphasizing larger geometric changes over smaller ones. Here we present DiffNets, self-supervised autoencoders that avoid such assumptions, and automatically identify the relevant features, by requiring that the low-dimensional representations they learn are sufficient to predict the biochemical differences between protein variants. For example, DiffNets automatically identify subtle structural signatures that predict the relative stabilities of β-lactamase variants and duty ratios of myosin isoforms. DiffNets should also be applicable to understanding other perturbations, such as ligand binding.
more »
« less
- Award ID(s):
- 2029322
- PAR ID:
- 10230219
- Publisher / Repository:
- Nature Publishing Group
- Date Published:
- Journal Name:
- Nature Communications
- Volume:
- 12
- Issue:
- 1
- ISSN:
- 2041-1723
- Format(s):
- Medium: X
- Sponsoring Org:
- National Science Foundation
More Like this
-
-
Abstract Structural, regulatory and enzymatic proteins interact with DNA to maintain a healthy and functional genome. Yet, our structural understanding of how proteins interact with DNA is limited. We present MELD-DNA, a novel computational approach to predict the structures of protein–DNA complexes. The method combines molecular dynamics simulations with general knowledge or experimental information through Bayesian inference. The physical model is sensitive to sequence-dependent properties and conformational changes required for binding, while information accelerates sampling of bound conformations. MELD-DNA can: (i) sample multiple binding modes; (ii) identify the preferred binding mode from the ensembles; and (iii) provide qualitative binding preferences between DNA sequences. We first assess performance on a dataset of 15 protein–DNA complexes and compare it with state-of-the-art methodologies. Furthermore, for three selected complexes, we show sequence dependence effects of binding in MELD predictions. We expect that the results presented herein, together with the freely available software, will impact structural biology (by complementing DNA structural databases) and molecular recognition (by bringing new insights into aspects governing protein–DNA interactions).more » « less
-
DNA damage and repair are widely studied in relation to cancer and therapeutics. Y-family DNA polymerases can bypass DNA lesions, which may result from external or internal DNA damaging agents, including some chemotherapy agents. Overexpression of Y-family polymerase human pol kappa can result in tumorigenesis and drug resistance in cancer. This report describes the use of computational tools to predict the effects of single nucleotide polymorphism variants on pol kappa activity. Partial Order Optimum Likelihood (POOL), a machine learning method that uses input features from Theoretical Microscopic Titration Curve Shapes (THEMATICS), was used to identify amino acid residues most likely involved in catalytic activity. The μ4 value, a metric obtained from POOL and THEMATICS that serves as a measure of the degree of coupling between one ionizable amino acid and its neighbors, was then used to identify which protein mutations are likely to impact the biochemical activity. Bioinformatic tools SIFT, PolyPhen-2, and FATHMM predicted most of these variants to be deleterious to function. Along with computational and bioinformatic predictions, we characterized the catalytic activity and stability of seventeen cancer-associated DNA pol kappa variants. We identified pol kappa variants R48I, H105Y, G147D, G154E, V177L, R298C, E362V, and R470C as having lower activity relative to wild-type pol kappa; the pol kappa variants T102A, H142Y, R175Q, E210K, Y221C, N330D, N338S, K353T, and L383F were identified as similar in catalytic efficiency to WT pol kappa. We observed that POOL predictions can be used to predict which variants have decreased activity. Predictions from bioinformatic tools like SIFT, PolyPhen-2 and FATHMM are based on sequence comparisons and therefore are complementary to POOL but are less capable of predicting biochemical activity. These bioinformatic and computational tools can be used to identify SNP variants with deleterious effects and altered biochemical activity from a large data set.more » « less
-
Abstract MotivationGiven a protein of unknown function, fast identification of similar protein structures from the Protein Data Bank (PDB) is a critical step for inferring its biological function. Such structural neighbors can provide evolutionary insights into protein conformation, interfaces and binding sites that are not detectable from sequence similarity. However, the computational cost of performing pairwise structural alignment against all structures in PDB is prohibitively expensive. Alignment-free approaches have been introduced to enable fast but coarse comparisons by representing each protein as a vector of structure features or fingerprints and only computing similarity between vectors. As a notable example, FragBag represents each protein by a ‘bag of fragments’, which is a vector of frequencies of contiguous short backbone fragments from a predetermined library. Despite being efficient, the accuracy of FragBag is unsatisfactory because its backbone fragment library may not be optimally constructed and long-range interacting patterns are omitted. ResultsHere we present a new approach to learning effective structural motif presentations using deep learning. We develop DeepFold, a deep convolutional neural network model to extract structural motif features of a protein structure. We demonstrate that DeepFold substantially outperforms FragBag on protein structural search on a non-redundant protein structure database and a set of newly released structures. Remarkably, DeepFold not only extracts meaningful backbone segments but also finds important long-range interacting motifs for structural comparison. We expect that DeepFold will provide new insights into the evolution and hierarchical organization of protein structural motifs. Availability and implementationhttps://github.com/largelymfs/DeepFoldmore » « less
-
Protein–DNA interactions play an important role in various biological processes such as gene expression, replication, and transcription. Understanding the important features that dictate the binding affinity of protein-DNA complexes and predicting their affinities is important for elucidating their recognition mechanisms. In this work, we have collected the experimental binding free energy (ΔG) for a set of 391 Protein-DNA complexes and derived several structure-based features such as interaction energy, contact potentials, volume and surface area of binding site residues, base step parameters of the DNA and contacts between different types of atoms. Our analysis on relationship between binding affinity and structural features revealed that the important factors mainly depend on the number of DNA strands as well as functional and structural classes of proteins. Specifically, binding site properties such as number of atom contacts between the DNA and protein, volume of protein binding sites and interaction-based features such as interaction energies and contact potentials are important to understand the binding affinity. Further, we developed multiple regression equations for predicting the binding affinity of protein-DNA complexes belonging to different structural and functional classes. Our method showed an average correlation and mean absolute error of 0.78 and 0.98 kcal/mol, respectively, between the experimental and predicted binding affinities on a jack-knife test. We have developed a webserver, PDA-PreD (Protein-DNA Binding affinity predictor), for predicting the affinity of protein-DNA complexes and it is freely available at https://web.iitm.ac.in/bioinfo2/pdapred/more » « less
