Jianzhu Ma

FoldAbS: Repurposing the Protein Folding Model as a Foundation Encoder for Antibody Screening

FoldAbS repurposes protein folding models as foundation encoders to improve antibody screening by evaluating specificity, affinity, and developability in a unified framework.

Programming Biomolecular Interactions with All-Atom Generative Model

This paper introduces AnewOmni, a generative model that designs biomolecules across scales by learning interaction patterns and physical constraints from 5 million complexes.

Atom-level generative foundation model for molecular interaction with pockets

This paper introduces PocketXMol, a generative foundation model for molecular interactions with protein pockets, achieving state-of-the-art performance in 11 tasks and demonstrating success in designing inhibitors and peptides.

The Human-In-the-Loop Drug Design Framework with Equivariant Rectified Flow

The paper introduces HIL-DD, a human-in-the-loop drug design framework that uses an Equivariant Rectified Flow Model to enable efficient collaboration between human experts and AI for generating drug-like molecules.

SpecLig: Energy-Guided Hierarchical Model for Target-Specific 3D Ligand Design

SpecLig generates target-specific ligands with high affinity and specificity by integrating hierarchical modeling and energy guidance, improving therapeutic potential and reducing off-target risks.

Peptide Design through Binding Interface Mimicry

PepMimic, an all-atom latent diffusion model guided by a contrastive interface encoder, designs peptides mimicking known binders; 8% of tested peptides across five targets reached 10^-8M KD.

Escaping the drug-bias trap: using debiasing design to improve interpretability and generalization of drug-target interaction prediction

This paper introduces UdanDTI, a deep-learning model for drug-target interaction prediction that addresses the drug-bias trap, showing improved accuracy and interpretability across datasets and EGFR mutation subgroups.

Pisces: A multi-modal data augmentation approach for drug combination synergy prediction

Pisces enhances drug combination synergy prediction by generating 64 augmented views per drug pair, achieving state-of-the-art results across multiple prediction tasks.

Rag2Mol: Structure-based drug design based on Retrieval Augmented Generation

This paper introduces Rag2Mol methods for structure-based drug design using retrieval-augmented generation, producing drug candidates with better binding affinities and drug-likeness than existing models.

FlowDesign: Improved Design of Antibody CDRs Through Flow Matching and Better Prior Distributions

FlowDesign improves antibody CDR design using flow matching and better priors, achieving higher affinity and neutralizing potency against HIV mutants.

Large Scale Foundation Model on Single-cell Transcriptomics

This paper introduces scFoundation, a large-scale pretrained model for single-cell transcriptomics, achieving state-of-the-art results in multiple downstream tasks.

Sagittarius: Extrapolating Heterogeneous Time-Series Gene Expression Data

Sagittarius simulates gene expression at unmeasured timepoints using a transformer model, improving understanding of temporal dynamics in biology.

xTrimoGene: An Efficient and Scalable Representation Learner for Single-Cell RNA-Seq Data

xTrimoGene is a scalable transformer model for scRNA-seq data that reduces computational demands while maintaining accuracy, achieving state-of-the-art results in downstream tasks.

Rotamer Density Estimator is an Unsupervised Learner of the Effect of Mutations on Protein-Protein Interaction

The paper introduces RDE, an unsupervised method that predicts mutation effects on protein-protein binding by estimating conformational flexibility from protein structures, outperforming existing methods.

High-resolution de novo structure prediction from primary sequence

From a single sequence, OmegaFold reaches mean LDDT 0.82 on 146 CAMEO targets, between MSA-based RoseTTAFold (0.75) and AlphaFold2 (0.86), and runs about ten times faster.

Proximal Exploration for Model-guided Protein Sequence Design

This paper introduces Proximal Exploration (PEX) and MuFacNet to improve protein sequence design by prioritizing low-mutation high-fitness mutants and predicting mutational effects.

scPretrain: Multi-task self-supervised learning for cell type classification

scPretrain improves cell type classification by using self-supervised learning with both annotated and unannotated cells across 60 datasets.

Antigen-Specific Antibody Design and Optimization with Diffusion-Based Generative Models for Protein Structures

This paper introduces a diffusion-based generative model for designing and optimizing antigen-specific antibodies by jointly modeling CDR sequences and structures.

Contrastive Cycle Adversarial Autoencoders for Single-cell Multi-omics Alignment and Integration

This paper introduces a framework to align and integrate single-cell RNA-seq and ATAC-seq data, achieving higher accuracy than existing methods on both simulated and real data.

Mapping cell structure across scales by fusing protein images and interactions

The paper integrates protein imaging and interaction data to create a unified map of eukaryotic cell structure, revealing new subcellular systems and protein associations.

cancerAlign: Stratifying tumors by unsupervised alignment across cancer types

cancerAlign improves tumor stratification by aligning tumors across cancer types using adversarial learning and consensus clustering, showing better results than methods that cluster within single cancer types.

Artificial-Cell-Type Aware Cell Type Classification in CITE-seq

CITE-sort improves cell type classification in CITE-seq by addressing artificial cell types caused by multiplets, achieving better clustering performance than existing methods.

Deep functional synthesis: a machine learning approach to gene functional enrichment

This paper introduces DeepSyn, a machine learning method for gene functional enrichment that uses a knowledge graph to infer gene functions based on disease and drug context, achieving high accuracy in various applications.

Mitigating Data Scarcity in Protein Binding Prediction Using Meta-Learning

The paper introduces a meta-learning framework to predict phosphorylation sites for kinases with limited experimental data, showing improved transferability and accuracy.

HiCluster: A Robust Single-Cell Hi-C Clustering Method Based on Convolution and Random Walk

HiCluster improves clustering of low coverage single-cell Hi-C data by imputing using linear convolution and random walk, revealing TAD-like structures consistent with bulk data.

Deciphering signaling specificity with interpretable deep neural networks

DeepSignal predicts kinase and SH2 domain substrate specificity using deep learning, outperforming existing methods and identifying cancer-related mutations and pathways.

DeepBound: Accurate Identification of Transcript Boundaries via Deep Convolutional Neural Fields

DeepBound identifies transcript boundaries from RNA-seq reads using deep convolutional neural fields, outperforming existing methods on simulated and biological datasets.