Yang Shen

ST-PARM improves multi-objective protein design by aligning a frozen language model with user-specified trade-offs using a reward model that handles uncertainty and achieves broader Pareto coverage.

This paper introduces an antibody language model (AbLM) to accelerate IgG screening and design for broad-spectrum antiviral therapy, achieving higher precision in predicting variant-resistant antibodies.

The paper evaluates variant effect predictions for ARSA gene variants, finding that a bootcamp-trained model outperformed others and that deep learning slightly improved performance over simpler methods.

The paper evaluates variant prioritization methods for rare disease diagnosis using data from the Rare Genomes Project, finding that top models identify causal variants in most solved families and uncover potential candidates in unsolved ones.

The paper evaluates 23 imputation models for cross-cell type epigenomic data, finding quality assessment challenging due to distributional shifts, data availability, and redundant metrics.

The paper develops a multimodal deep learning model that combines 1D genome sequence and 3D chromatin structure to predict noncoding variant effects, outperforming sequence-only models.

The paper investigates whether multi-modal data and auxiliary tasks improve inter-protein contact prediction, finding that multi-modal inputs enhance precision but auxiliary tasks have limited benefits.

This paper introduces cross-modality and self-supervised learning for protein embeddings to improve compound-protein affinity and contact prediction.

TALE is a deep learning model for protein function annotation that uses joint sequence-label embeddings to improve generalizability to novel sequences, species, and functions.

This paper develops interpretable deep learning models for structure-free compound-protein affinity prediction and contact mapping, showing improved interpretability and performance.

This study predicts compound-protein affinity and contact using cross-modality protein embeddings, showing improved performance over single-modality approaches.

This paper develops a deep generative model, gcWGAN, to design novel protein folds using sequence-structure data, achieving higher fold coverage and sequence diversity than a competing method.

This paper introduces a Bayesian Active Learning algorithm for protein docking that provides uncertainty quantification and improves prediction accuracy.

This paper introduces a deep generative model for designing drug combinations using graph-based principles, showing it can generate low-toxicity combinations that match known drug modules.

This paper introduces EGCN, an energy-based graph convolutional network for protein docking that improves model ranking and quality assessment.

This paper introduces a weakly supervised regression model to predict pathogenicity of missense variants and infer molecular mechanisms from imprecise annotations.

DeepAffinity predicts compound-protein affinity using a semi-supervised deep learning model combining recurrent and convolutional networks, achieving high accuracy and interpretability.