Abstract
Isoforms from the same gene can significantly rewire protein interaction networks, but proteome-wide computational evaluation of these effects remains challenging. In this work, we present DeepISO, a deep learning framework for predicting isoform-specific interactions. DeepISO integrates two graph convolutional neural networks and a random forest model via a logistic regression model. To the best of our knowledge, this is the first approach to jointly leverage AlphaFold-predicted structures and ESM2 language model embeddings for this task. Compared with state-of-the-art PPI prediction tools, DeepISO demonstrates superior performance.