SLSa Li

Sa Li

*This is an algorithmically generated author record based on publicly available information.

Current Organizations / Institutes

The organizations/institutes from the scholar's latest article


Sa Li has a total of 4 co-authors. Here are the top co-authors by number of shared publications:


Latest Publications

  • Open Access

    Learning Interpretable Gene Representations with Adaptive Chebyshev Graph Neural Networks

    by Li Sa, Tianle Ma

    Challenging due to the heterogeneity of molecular modalities and the complex structure of gene interaction networks. Effective integration of omics data with protein–protein interaction networks requires models that capture multiscale dependencies while remaining interpretable. To address these challenges, we propose ACGNN, an adaptive Chebyshev graph neural network for cancer driver gene prediction.ACGNN integrates pan-cancer multi-omics features with protein–protein interaction networks using adaptive Chebyshev spectral filtering to model multiscale graph structure. The framework employs a two-stage learning strategy: an auxiliary-task pretraining stage to learn modality-specific gene embeddings from omics- and topology-derived signals, followed by task-specific fine-tuning for driver gene prediction. To support model interpretation, ACGNN incorporates pathway-level analysis, omics-level attribution using Integrated Gradients, and spectral biclustering to identify biologically coherent gene modules.Performance evaluation on benchmark datasets shows that ACGNN improves driver gene prediction compared with existing graph-based methods, while yielding biologically meaningful pathway associations and omics-level attributions. These results indicate that adaptive spectral graph learning supports accurate and interpretable identification of cancer driver genes.

  • Open Access

    Pathway Representation Learning for Interpretable Graph Neural Networks

    by Sa Li, Tianle Ma

    Pathway-level modeling offers a biologically grounded abstraction for molecular prediction, yet existing computational approaches exhibit important limitations. Traditional pathway analysis methods rely on independent enrichment statistics and ignore pathway-pathway interactions, while recent graph neural network (GNN) models predominantly operate at the gene or protein level. These gene-centric GNNs often require high-dimensional molecular inputs, incur substantial computational cost, and provide limited interpretability. Moreover, uniform neighborhood aggregation treats all pathways equivalently, obscuring pathway-specific heterogeneity and limiting mechanistic insight, particularly in pathway-pathway inference tasks. We present SpecPathGNN, a spectral graph learning framework that elevates pathway enrichment from a preprocessing step to a core representational principle. Molecular measurements are projected into a structured pathway space via enrichment analysis, where statistical significance is explicitly encoded as weighted edges and binary indicators. This construction yields a hierarchical pathway knowledge graph that preserves biological organization while substantially reducing dimensional complexity. SpecPathGNN adopts a two-stage learning paradigm. In the first stage, spectral pretraining learns pathway embeddings that respect global graph topology and hierarchical structure. In the second stage, these embeddings are fine-tuned for downstream prediction through a GNN-based encoder and a multilayer perceptron decoder, enabling accurate pathway-pathway link prediction. Beyond performance, SpecPathGNN is designed for interpretability. By integrating attention-based attribution, gradient-driven node saliency, and saliency-guided spectral biclustering, the framework uncovers coherent gene-pathway modules supported by statistically meaningful enrichment signals. These modules provide mechanistic explanations that align with known biology while revealing novel pathway interactions. Extensive experiments demonstrate that SpecPathGNN consistently outperforms the state-of-the-art GNN baselines in embedding quality and link prediction accuracy. All data and source code are publicly available at https://github.com/Ericsaliway0/SpecPathGNN.


  • Open Access

    SGL-CDP: Spectral Graph Learning for Cancer Driver Prioritization

    by Sa Li, Tianle Ma

    Identifying cancer driver genes (CDGs) remains a fundamental challenge in cancer genomics, as effective prioritization requires integrating heterogeneous multi-omics signals while modeling the complex topology of gene–gene interaction networks. Existing computational approaches, including many graph neural network (GNN) models, often rely on a single interaction network or employ shallow feature fusion strategies, limiting their ability to capture multiscale molecular dependencies underlying tumorigenesis. Moreover, limited interpretability restricts their translational utility in biomedical research.We propose SGL-CDP (Spectral Graph Learning for Cancer Driver Prioritization), a spectral graph learning framework that jointly models multi-omics features and network topology within a unified architecture. SGL-CDP adopts a two-stage training paradigm. In the pretraining stage, spectral representations are learned independently for each omics modality–including gene expression, DNA methylation, copy number alteration, and somatic mutation–as well as for graph topology, enabling the model to capture modality-specific molecular and structural characteristics. In the fine-tuning stage, these pretrained representations are integrated and jointly optimized using curated cancer driver annotations to improve prioritization accuracy. This design preserves biologically meaningful embeddings, mitigates overfitting, and enhances cross-modality alignment. Across multiple tumor cohorts, SGL-CDP achieves strong predictive performance and reveals biologically coherent pathway-level organization through interpretable cluster–feature associations. The results demonstrate that adaptive spectral graph learning provides an effective and interpretable framework for cancer driver gene prioritization. Data and source code are available at: https://github.com/Ericsaliway0/SGL-CDP.git.

  • Integration of Multi-Omics Data With Topology Adaptive Graph Convolutional Network for Cancer Driver Gene Identification

    by Sa Li, Jonah Shader, Abhijeet Bhattacharya, Tianle Ma

    The increasing volume of high-throughput molecular data has presented substantial computational challenges in the identification of cancer driver genes. We introduce ATTAG, a framework based on topology-adaptive graph neural networks with an attention mechanism, aimed at predicting cancer driver genes by integrating multi-omics data with biomolecular networks. ATTAG creates three distinct gene networks derived from protein interactions, gene semantic similarities, and co-occurrence within signaling pathways. Graph convolutional networks (GCNs) are employed to generate low-dimensional embeddings from these networks and multi-omics data. These embeddings are then refined through topology-adaptive graph neural networks for the prediction of cancer driver genes in BLCA. When compared to state-of-the-art methods, ATTAG demonstrates exceptional performance in identifying cancer driver genes. The experimental results underscore its ability to accurately identify both well-known driver genes and novel candidate cancer genes.