Mirko Torrisi

This study evaluates if pre-training chemical language models on large compound datasets improves compound representation, finding that standard masked language modeling is insufficient.

This paper improves the assessment of deep learning models for drug-target interaction prediction by addressing dataset biases and proposing a more accurate evaluation framework.

The paper evaluates the role of evolutionary profiles in protein secondary structure prediction and introduces a new protocol to fairly compare profile-based predictors.

Porter 5 predicts ab initio protein secondary structure at 84% 3-class accuracy and 73% 8-class accuracy on an independent test set, available as web and standalone tools.