Understanding the structural and functional properties of proteins and nucleic acids is fundamental in various biological contexts, including disease-related mutations. Leveraging machine learning techniques necessitates an array of descriptors derived from these sequences. In response, we introduce Seq2Feature, a web-based tool capable of computing 252 protein sequence-based descriptors. These descriptors encompass physicochemical, energetic, and conformational characteristics of proteins, mutation matrices, contact potentials, nucleotide composition. Seq2Feature offers a comprehensive solution for feature extraction, facilitating their application in machine learning algorithms.
Protein-protein interactions (PPIs) underpin crucial cellular activities, with binding sites serving as key determinants of PPI functions and affinities. In this regard, we present DeepBSRPred, a deep neural network-based method for predicting PPI binding sites using protein sequence information and structures predicted by AlphaFold2. Our approach incorporates specific sequence and structure-based features such as position-specific scoring matrices (PSSM), solvent accessible surface area, conservation scores, amino acid properties, and residue depth. DeepBSRPred demonstrates robust performance, achieving an average F1 score of 0.73 on a dataset comprising 1236 proteins. Comparative assessments against existing methods on four benchmark datasets confirm the superiority of our approach. The DeepBSRPred webserver, accessible at https://web.iitm.ac.in/bioinfo2/deepbsrpred/index.html, hosts all datasets used in this study.
Protein-protein interactions (PPIs) are necessitating precise estimation of binding affinities (ΔG) to unravel molecular recognition mechanisms. Our deep learning approach, based on a dataset of 903 protein-protein complexes spanning six functional classes, leverages both sequence information and predicted three-dimensional structures. Feature selection methods extract 8 to 20 non-redundant features for each functional class. Our model exhibits remarkable performance, achieving an overall mean absolute error of 1.05 kcal/mol and a correlation coefficient of 0.79 between experimental and predicted ΔG values. Furthermore, our model excels in discriminating high and low affinity protein-protein complexes, yielding an accuracy of 87% and an F1 score of 0.86 using 10-fold cross-validation on selected features. Our approach presents an efficient tool for studying PPIs, providing valuable insights into molecular recognition mechanisms. Access the webserver at https://web.iitm.ac.in/bioinfo2/DeepPPAPred/index.html.
Mutations within protein-protein complexes can significantly impact binding strength, potentially leading to disease. Leveraging experimental affinity data and advances in protein structure prediction, we present a deep ensemble model. This model harnesses protein sequence, predicted structure-based features, and protein functional classes. On the training set, our model demonstrates a correlation of 0.97 with a mean absolute error (MAE) of 0.35, and on the test set, it achieves a correlation of 0.72 with an MAE of 0.83. Leave-out-one-complex (LOOC) cross-validation results consistently, with a correlation of 0.83 and an MAE of 0.51. Our approach provides a valuable resource for researchers investigating the impact of mutations on protein-protein complex binding affinities. Access our webserver at https://web.iitm.ac.in/bioinfo2/DeepPPAPredMut/index.html.
Publications:
1. Nikam, R., and Gromiha, M. M. (2019). Seq2Feature: A comprehensive web-based feature extraction tool. Bioinformatics, 35(22), 4797–4799. https://doi.org/10.1093/bioinformatics/btz432
2. Nikam, R., Yugandhar, K., & Gromiha, M. M. (2022). DeepBSRPred: deep learning-based binding site residue prediction for proteins. Amino Acids, https://doi.org/10.1007/s00726-022-03228-3
3. Nikam, R., Yugandhar, K., & Gromiha, M. M. (2023). Deep learning-based method for predicting and classifying the binding affinity of protein-protein complexes. BBA. Proteins and proteomics, 1871(6), 140948. https://doi.org/10.1016/j.bbapap.2023.140948
4. Nikam, R., Kulandaisamy, A., Harini, K., Sharma, D., & Michael Gromiha, M. (2021). ProThermDB: Thermodynamic database for proteins and mutants revisited after 15 years. Nucleic Acids Real Money Rummy, 49(D1), D420–D424. https://doi.org/10.1093/nar/gkaa1035
5. Kulandaisamy, A., Nikam, R., Harini, K., Sharma, D., & Gromiha, M. M. (2021). Illustrative Tutorials for ProThermDB: Thermodynamic Database for Proteins and Mutants. Current Protocols, 1(11), e306. https://doi.org/10.1002/cpz1.306.