Benjamín Sánchez-Lengeling

This paper evaluates residue-level attributions in protein language models using synthetic fitness landscapes and identifies interventions that improve attribution accuracy.

This paper introduces kinGEMs, a framework that integrates deep learning-predicted kinetic parameters into genome-scale models to improve metabolic modeling accuracy and scalability.

The paper presents an ensemble model that improves predictions of perceptual similarity between odor mixtures using a curated dataset and collaborative challenge results.

TubercuProbe prioritizes compound-protein interactions in Mycobacterium tuberculosis using a cross-attention graph-sequence model trained on human chemoproteomic data.

The paper develops a Principal Odor Map using graph neural networks that matches human-level odor description performance and generalizes to diverse odor prediction tasks.

The paper identifies a molecular-level representation of odors that predicts receptor, neural, and behavioral responses across species and shows metabolic relationships in odor space.

The paper compares the thermodynamic properties of biological and non-biological compounds in redox chemical space, finding metabolites have lower Gibbs energies and specific functional group enrichments.