Xiaoyong Pan

The paper evaluates models for classifying synonymous variants by separating gene context, exon position, and variant-specific effects, finding that gene identity and position are more predictive than the variant itself.

AuditPPI evaluates PPI predictors by transforming language model features into auditable fingerprints, revealing that predictive success does not necessarily reflect biological signal capture.

IgGM2 is an all-atom foundation model for immune receptor design that jointly predicts CDR sequences and full-atom structures, outperforming existing methods in structure prediction and sequence design benchmarks.

RNAJog optimizes RNA sequences for multiple objectives using autoregressive reinforcement learning, achieving faster performance and improved vaccine design outcomes.

This paper introduces AlignIF, a method for 3D RNA inverse folding that uses structural alignments and a novel encoder to generate RNA sequences with improved performance.

RBPSignal predicts RNA-protein binding signals using deep learning and identifies sequence motifs, validated against eCLIP data and existing databases.

This study introduces NLSExplorer, a deep learning model for identifying nuclear localization signals, showing improved performance on benchmark datasets and uncovering NLS patterns across 416 species.

This study introduces DGP-AMIO, a graph attention network-based method that integrates multi-omics and gene interaction data to identify novel disease genes, achieving high performance across 20 disease datasets.

DGRNA is a long-context RNA foundation model using bidirectional Mamba2 that outperforms existing RNA language models on six downstream tasks.

stFormer is a transformer model for spatial transcriptomics data that improves performance on cell clustering and gene function prediction compared to existing models.

GexMolGen generates hit-like molecules using gene expression signatures through a cross-modal approach, producing valid molecules from both in- and out-of-domain transcriptome data.

AttSiOff improves siRNA design by predicting inhibition and off-target effects using self-attention mechanisms, achieving better accuracy and generalization than existing methods.

The paper introduces CircSite, a deep learning tool that predicts RBP binding sites on circRNAs at single-nucleotide resolution and identifies key binding subsequences.

TFvelo improves RNA velocity estimation by incorporating gene regulatory information, achieving accurate dynamics modeling without relying on splicing data.

DeepSipred predicts siRNA inhibition using sequence, thermodynamic, and expert knowledge features, achieving state-of-the-art performance on public datasets.

This paper introduces EGNA, an empirical graph neural network for predicting protein-ligand binding affinity, showing improved accuracy over existing methods on benchmark datasets.

OUGene 2.0 is an updated database of disease-associated over- and under-expressed genes, mined from full-text articles, containing 197,236 associations.

The paper introduces CRIP, a deep learning method for predicting circRNA-RBP interactions using RNA sequences, outperforming existing methods with a hybrid CNN-RNN architecture.

This study introduces DimiG, a semi-supervised multi-label graph convolutional network to infer disease-miRNA associations, showing improved performance over existing methods.

This study introduces node2loc, a deep learning method that predicts protein subcellular locations using protein-protein network representations, achieving an MCC of 0.812.

This study introduces iDeepS, a deep learning method that predicts RNA-protein binding sites by integrating sequence and structure information, outperforming existing methods.

The paper introduces iDeep, a hybrid deep learning framework that improves RNA-protein binding motif prediction by integrating cross-domain knowledge, achieving better performance and interpretability.