Shuangjia Zheng

AlphaEnzyme improves enzyme discovery by predicting catalytic properties and prioritizing sequence changes, achieving higher success rates in catalyst retrieval and fitness enhancement.

The paper benchmarks 13 models for predicting single-cell perturbation responses across 25 datasets, finding performance varies with perturbation effect size and evaluation metric.

The paper introduces a framework for self-auditing AI in drug discovery, using verifiers to validate predictions and adjust objectives based on experimental results.

VeloTrace unifies RNA velocity and trajectory inference using NeuralODEs, improving accuracy and coherence in single-cell transcriptomic analysis.

ProTact improves protein-protein interaction prediction using a geometric graph neural network with physics-informed constraints, outperforming existing methods on benchmark datasets.

EVA is a long-context generative RNA model trained on 114 million sequences, achieving state-of-the-art performance in RNA design tasks and structural modeling.

ECloudGen improves molecular design by using electron clouds as latent variables, generating more potent binders and exploring broader chemical space than existing methods.

OriGene identifies therapeutic targets by integrating diverse data modalities and outperforms human experts and models in accuracy and robustness.

FoldBench evaluates all-atom biomolecular structure prediction models across nine tasks, revealing performance gaps and highlighting AlphaFold 3's superiority.

The paper introduces a zero-shot RNA design method using genomic foundation models, achieving improved fluorescence and binding affinity in RNA aptamers.

GerNA-Bind predicts RNA-ligand binding specificity using a geometric deep learning framework, achieving improved precision and identifying compounds with experimentally confirmed activity against MALAT1 RNA.

EnzymeCAGE is a geometric foundation model trained on enzyme-reaction pairs to predict enzyme function, de-orphan reactions, identify catalytic sites, and reconstruct biosynthetic pathways.

The paper curates a large dataset of protein-protein interaction mutations to improve deep learning models for predicting interaction strengths and applications in drug discovery.

The paper evaluates computational hit-finding methods for targeting the WDR domain of LRRK2, a Parkinson's disease protein, finding that diverse strategies yielded weakly potent compounds.

The paper proposes a cross-modal graph contrastive learning method that uses cellular images to improve molecular representation learning for downstream tasks.

This paper benchmarks AlphaFold3 against SKEMPI to assess its ability to predict mutational effects on protein-protein interactions and compares it with force field and deep learning methods.

This paper introduces GraphEBM, an SE(3)-equivariant graph attention network for protein side chain conformation, outperforming existing methods in rotamer recovery and combinatorial optimization.

The paper introduces GraphSite, a method using AlphaFold2-predicted structures and graph transformers to predict DNA-binding residues, outperforming existing sequence and structure-based methods.

DeepANIS predicts antibody paratopes by integrating BiLSTM and transformer networks on concatenated CDR sequences, outperforming existing methods.

X-DPI is a structure-aware multi-modal deep learning model for drug-protein interactions prediction that outperforms existing methods on benchmark datasets.

The paper develops a structure-aware model using graph convolutional networks and predicted contact maps to predict protein solubility from sequence, showing improved performance over existing methods.

The paper develops MolRep, a library unifying 16 molecular representation models, and evaluates them on 12 benchmarks to identify top-performing models for property prediction.

This paper predicts drug-protein interactions using a deep learning framework that combines 2D protein distance maps and drug molecular notation, achieving competitive performance on benchmark datasets.

The paper introduces MTDsite, a multi-task deep learning method for predicting protein-binding residues across DNA, RNA, peptide, and carbohydrate, achieving higher accuracy than existing methods.