Zinnia Ma

This paper shows that protein language models can detect local structural similarities from sequence data without 3D models, identifying aligned regions through cosine similarity of residue embeddings.

This paper introduces RemoteFoldSet and the Structural Awareness score to benchmark how well protein language models encode structural information.

This paper introduces TopoDiff, a diffusion-based framework for protein backbone generation that improves global geometry control and fold diversity through a structure encoder and latent sampler.

DrugChat, a multi-modal LLM, predicts drug mechanisms and properties through free-form text, outperforming GPT-4 and specialized models in accuracy and versatility.