Ensemble Learning with Supervised Methods Based on Large-Scale Protein Language Models for Protein Mutation Effects Prediction

Machine learning has been increasingly utilized in the field of protein engineering, and research directed at predicting the effects of protein mutations has attracted increasing attention. Among them, so far, the best results have been achieved by related methods based on protein language models, which are trained on a large number of unlabeled protein sequences to capture the generally hidden evolutionary rules in protein sequences, and are therefore able to predict their fitness from protein sequences. Although numerous similar models and methods have been successfully employed in practical protein engineering processes, the majority of the studies have been limited to how to construct more complex language models to capture richer protein sequence feature information and utilize this feature information for unsupervised protein fitness prediction. There remains considerable untapped potential in these developed models, such as whether the prediction performance can be further improved by integrating different models to further improve the accuracy of prediction. Furthermore, how to utilize large-scale models for prediction methods of mutational effects on quantifiable properties of proteins due to the nonlinear relationship between protein fitness and the quantification of specific functionalities has yet to be explored thoroughly. In this study, we propose an ensemble learning approach for predicting mutational effects of proteins integrating protein sequence features extracted from multiple large protein language models, as well as evolutionarily coupled features extracted in homologous sequences, while comparing the differences between linear regression and deep learning models in mapping these features to quantifiable functional changes. We tested our approach on a dataset of 17 protein deep mutation scans and indicated that the integrated approach together with linear regression enables the models to have higher prediction accuracy and generalization. Moreover, we further illustrated the reliability of the integrated approach by exploring the differences in the predictive performance of the models across species and protein sequence lengths, as well as by visualizing clustering of ensemble and non-ensemble features.

Protein Fitness Prediction Is Impacted by the Interplay of Language Models, Ensemble Learning, and Sampling Methods

Advances in machine learning (ML) and the availability of protein sequences via high-throughput sequencing techniques have transformed the ability to design novel diagnostic and therapeutic proteins. ML allows protein engineers to capture complex trends hidden within protein sequences that would otherwise be difficult to identify in the context of the immense and rugged protein fitness landscape. Despite this potential, there persists a need for guidance during the training and evaluation of ML methods over sequencing data. Two key challenges for training discriminative models and evaluating their performance include handling severely imbalanced datasets (e.g., few high-fitness proteins among an abundance of non-functional proteins) and selecting appropriate protein sequence representations (numerical encodings). Here, we present a framework for applying ML over assay-labeled datasets to elucidate the capacity of sampling techniques and protein encoding methods to improve binding affinity and thermal stability prediction tasks. For protein sequence representations, we incorporate two widely used methods (One-Hot encoding and physiochemical encoding) and two language-based methods (next-token prediction, UniRep; masked-token prediction, ESM). Elaboration on performance is provided over protein fitness, protein size, and sampling techniques. In addition, an ensemble of protein representation methods is generated to discover the contribution of distinct representations and improve the final prediction score. We then implement multiple criteria decision analysis (MCDA; TOPSIS with entropy weighting), using multiple metrics well-suited for imbalanced data, to ensure statistical rigor in ranking our methods. Within the context of these datasets, the synthetic minority oversampling technique (SMOTE) outperformed undersampling while encoding sequences with One-Hot, UniRep, and ESM representations. Moreover, ensemble learning increased the predictive performance of the affinity-based dataset by 4% compared to the best single-encoding candidate (F1-score = 97%), while ESM alone was rigorous enough in stability prediction (F1-score = 92%).