Jianyang Zeng

ProMeta improves disease diagnosis and prediction from plasma proteomics using a meta-learning framework that overcomes data scarcity with biological priors and interpretable features.

SpaceTracer detects somatic SNVs from spatial transcriptomics data to enable lineage tracing in human tissues, revealing tumor progression and immune cell migration.

This paper introduces DeepSCFold, a method for improving protein complex structure prediction by using sequence-based deep learning to model inter-chain interactions, achieving higher accuracy than existing methods.

STARNet infers gene regulatory networks from spatial multi-omics data, revealing domain-specific regulatory interactions in tissues.

This paper introduces scPRS, a deep learning model that deconvolves polygenic risk scores at the single-cell level to uncover cellular and molecular disease mechanisms in complex human diseases.

The paper identifies CCDC146 as a modifier of ALS survival through primary cilia dysfunction and shows ASO targeting of CCDC146 rescues ALS phenotypes in human and mouse models.

PRS-Net improves disease prediction by modeling genetic nonlinearity using deep learning and identifies key genes and interactions influencing Alzheimer's and multiple sclerosis risk.

This paper presents a data-driven drug repositioning framework that identifies CVL218, a PARP1 inhibitor, as a potential therapeutic for COVID-19 through in silico and experimental validation.

This study introduces PULSE, a deep learning model that predicts pseudouridylation sites and characterizes their sequence features using convolutional neural networks.

MoTSE estimates task similarity for molecular property prediction to improve transfer learning performance by leveraging graph neural networks and interpretability.

CAMP predicts peptide-protein interactions and binding residues using a convolutional attention-based neural network, outperforming existing methods on benchmark datasets.

MONN predicts non-covalent interactions and binding affinities between compounds and proteins, outperforming existing methods and providing interpretable insights.

ACME improves peptide-MHC class I binding prediction using attention-based deep learning, outperforming existing methods and providing interpretable insights.

This paper introduces GEM-FISH, a method that integrates Hi-C and FISH data to improve 3D chromosome modeling, showing better accuracy in TADs, compartments, and super-enhancer distributions than Hi-C alone.

This study introduces RiboRL, a deep reinforcement learning framework for predicting ribosome densities on transcripts, showing improved performance over existing methods.

DeepHINT predicts HIV-1 integration sites and explains their selection using an attention-based deep learning framework, outperforming conventional methods.

Opal improves metagenomic binning by using low-density hashing to achieve faster and more accurate identification of sequencing read origins compared to existing methods.

NeoDTI predicts drug-target interactions by integrating heterogeneous network data, showing improved performance and novel predictions supported by prior studies.

GEM reconstructs 3D chromatin structures from Hi-C data using manifold learning and conformational energy, outperforming existing methods and recovering long-range interactions.

This paper introduces DTINet, a network integration method for predicting drug-target interactions and repositioning drugs, showing improved performance and validated novel COX inhibitor applications.

DeBooster improves RBP target prediction from CLIP-seq data by reducing false positives and negatives, and enables new biological insights into RBP functions.

TITER predicts translation initiation sites using deep learning on QTI-seq data, outperforming existing methods and identifying sequence signatures for different TIS codons.

ROSE predicts ribosome stalling from ribosome profiling data, showing better performance than conventional models and revealing regulatory factors and functional interactions.

Compound and protein embeddings learned from unlabelled data feed a deep network for compound-protein interaction prediction, reported to perform better on ChEMBL, BindingDB and DrugBank.