Search

As the application of large language models (LLMs) has broadened into the realm of biological predictions, leveraging their capacity for self-supervised learning to create feature representations of amino acid sequences, these models have set a new benchmark in tackling downstream challenges, such as subcellular localization. However, previous studies have primarily focused on either the structural design of models or differing strategies for fine-tuning, largely overlooking investigations into the nature of the features derived from LLMs. In this research, we propose different ESM2 representation extraction strategies, considering both the character type and position within the ESM2 input sequence. Using model dimensionality reduction, predictive analysis and interpretability techniques, we have illuminated potential associations between diverse feature types and specific subcellular localizations. Particularly, the prediction of Mitochondrion and Golgi apparatus prefer segments feature closer to the N-terminal, and phosphorylation site-based features could mirror phosphorylation properties. We also evaluate the prediction performance and interpretability robustness of Random Forest and Deep Neural Networks with varied feature inputs. This work offers novel insights into maximizing LLMs' utility, understanding their mechanisms, and extracting biological domain knowledge. Furthermore, we have made the code, feature extraction API, and all relevant materials available at https://github.com/yujuan-zhang/feature-representation-for-LLMs. [ABSTRACT FROM AUTHOR]

Throughout evolution, pathogenic viruses have developed different strategies to evade the response of the adaptive immune system. To carry out successful replication, some pathogenic viruses encode different proteins that manipulate the molecular mechanisms of host cells. Currently, there are different bioinformatics tools for virus research; however, none of them focus on predicting viral proteins that evade the adaptive system. In this work, we have developed a novel tool based on machine and deep learning for predicting this type of viral protein named VirusHound-I. This tool is based on a model developed with the multilayer perceptron algorithm using the dipeptide composition molecular descriptor. In this study, we have also demonstrated the robustness of our strategy for data augmentation of the positive dataset based on generative antagonistic networks. During the 10-fold cross-validation step in the training dataset, the predictive model showed 0.947 accuracy, 0.994 precision, 0.943 F1 score, 0.995 specificity, 0.896 sensitivity, 0.894 kappa, 0.898 Matthew's correlation coefficient and 0.989 AUC. On the other hand, during the testing step, the model showed 0.964 accuracy, 1.0 precision, 0.967 F1 score, 1.0 specificity, 0.936 sensitivity, 0.929 kappa, 0.931 Matthew's correlation coefficient and 1.0 AUC. Taking this model into account, we have developed a tool called VirusHound-I that makes it possible to predict viral proteins that evade the host's adaptive immune system. We believe that VirusHound-I can be very useful in accelerating studies on the molecular mechanisms of evasion of pathogenic viruses, as well as in the discovery of therapeutic targets. [ABSTRACT FROM AUTHOR]