DeProViR:a deep-learning framework based on pre-trained sequence embeddings for predicting host-viral protein-protein interactions


Emerging diseases like COVID-19 pose dual threats to public health and the economy. Understanding protein-protein interactions (PPIs) between viral and host proteins is crucial for antiviral therapies and studying pathogen replication. However, experimental techniques have limitations and machine learning models primarily focus on sequence-derived features, neglecting semantic information and necessitating effective encoding schemes. To address these challenges, we present DeProViR, a deep-learning framework for predicting virus-human interactions using amino acid sequences. DeProViR incorporates a Siamese-like neural network that combines convolutional and bidirectional LSTM networks to capture contextual information. Using GloVe embeddings, DeProViR seamlessly integrates semantic associations, enhancing PPI prediction. This innovative framework overcomes limitations in feature engineering and encoding scheme dependence. DeProViR provides an efficient solution for predicting host-virus interactions, facilitating therapy development, and advancing our understanding of diseases.

Presented at 2023 Conference