Barthelemy Meynard-Piganeau

Balancing Locality and Reconstruction in Protein Structure Tokenizer

This paper introduces a VQ-VAE-based protein structure tokenizer that balances reconstruction and retrieval, improving alignment with sequence models for better structure prediction.

TULIP - a Transformer based Unsupervised Language model for Interacting Peptides and T-cell receptors that generalizes to unseen epitopes

TULIP is an unsupervised transformer model for predicting TCR epitope binding, outperforming existing methods on unseen epitopes by addressing data limitations and bias.

Generating Interacting Protein Sequences using Domain-to-Domain Translation

The paper develops a method to generate protein domain sequences that interact with another domain, outperforming existing approaches and exploring fine-tuning and structure prediction.

Interpretable Pairwise Distillations for Generative Protein Sequence Models

The paper evaluates three neural network models for protein sequences, showing that pairwise models can replicate their performance in mutational effect prediction.