Gene selection in high-dimensional genomic data is essential for understanding disease mechanisms and improving therapeutic outcomes. Traditional feature selection methods effectively identify predictive genes but often ignore complex biological pathways and regulatory networks, leading to unstable and biologically irrelevant signatures. Prior approaches, such as Lasso-based methods and statistical filtering, either focus solely on individual gene-outcome associations or fail to capture pathway-level interactions, presenting a key challenge: how to integrate biological pathway knowledge while maintaining statistical rigor in gene selection? To address this gap, we propose a novel two-stage framework that integrates statistical selection with biological pathway knowledge using multi-agent reinforcement learning (MARL). First, we introduce a pathway-guided pre-filtering strategy that leverages multiple statistical methods alongside KEGG pathway information for initial dimensionality reduction. Next, for refined selection, we model genes as collaborative agents in a MARL framework, where each agent optimizes both predictive power and biological relevance. Our framework incorporates pathway knowledge through Graph Neural Network-based state representations, a reward mechanism combining prediction performance with gene centrality and pathway coverage, and collaborative learning strategies using shared memory and a centralized critic component. Extensive experiments on multiple gene expression datasets demonstrate that our approach significantly improves both prediction accuracy and biological interpretability compared to traditional methods.

Genomic datasets are often characterized by high dimensionality and a limited number of samples, making it difficult to identify biologically meaningful features. Effective feature selection is crucial in biomedical research, where both genomic and clinical variables are abundant. In this study, we evaluate multiple feature selection techniques on large-scale gene expression datasets from breast cancer, lung adenocarcinoma, and ovarian cancer. We specifically adapt and optimize a novel reinforcement learning (RL) approach to address the challenges of high-dimensional genomic data. The performance of the RL-based method is compared against widely-used techniques, including Least Absolute Shrinkage and Selection Operator (LASSO), Recursive Feature Elimination (RFE), Random Forest, and Principal Component Analysis (PCA). Our experimental results demonstrate that the RL approach outperforms traditional methods in predicting cancer outcomes in most cases, highlighting its potential for identifying biologically meaningful features in genomic analysis.

Creating an effective representation space is crucial for mitigating the curse of dimensionality, enhancing model generalization, addressing data sparsity, and leveraging classical models more effectively. Recent advancements in automated feature engineering (AutoFE) have made significant progress in addressing various challenges associated with representation learning, issues such as heavy reliance on intensive labor and empirical experiences, lack of explainable explic-itness, and inflexible feature space reconstruction embedded into downstream tasks. However, these approaches are constrained by: 1) generation of potentially unintelligible and illogical reconstructed feature spaces, stemming from the neglect of expert-level cognitive processes; 2) lack of systematic exploration, which subsequently results in slower model convergence for identification of optimal feature space. To address these, we introduce an interaction-aware reinforced generation perspective. We redefine feature space reconstruction as a nested process of creating meaningful features and controlling feature set size through selection. We develop a hierarchical reinforcement learning structure with cascading Markov Decision Processes to automate feature and operation selection, as well as feature crossing. By incorporating statistical measures, we reward agents based on the interaction strength between selected features, resulting in intelligent and efficient exploration of the feature space that emulates human decision-making. Extensive experiments are conducted to validate our proposed approach.

Multivariate time series data is prone to a high missing rate which presents an obstacle to statistical analysis of the data. Imputation has become the standard measure to handle this challenge. However, existing time series missing value imputation methods are mostly uni-modal that relies on self- imputation. With an unprecedented rate of data collection, the availability of multi-modal data is increasing, allowing us the opportunity to impute the time series missing values using other datasets generated from the same cohort. In this paper, we propose a multi-modal time series missing value imputation framework, TSEst, that can utilize multiple data modalities to overcome the limitations of self-imputation. The framework uses additional cross-sectional or time series data for the imputation and therefore, is less affected by a high missing rate in the time series data. A comprehensive set of experiments on two datasets shows an improvement in imputation accuracy over the baselines. Experimental results also demonstrate that the improvement is caused by the effective integration of the additional data modality. The proposed framework can impute missing values in the samples with no time series data available, reducing the reliance on long-term data collection. Availability: Code is available at https://github.com/compbiolabucf/TSEst

Effective feature selection can help reduce dimensionality, improve prediction accuracy, and increase result comprehensibility. Classic feature selection methods typically select and test feature subset in multiple iterations, and thus can be regarded as an exploratory process. In recent literature, the multi-armed bandit has become an emerging method to automate exploration for searching optimal solutions in large spaces. In this paper, our research question is: Can the multi-armed bandit formulation help us to automate feature selection? Along this line, we reformulate the feature selection problem with the combinatorial multi-armed bandit (CMAB) framework by regarding each feature as an arm. We propose two novel oracles and investigate how the super arm is formed under different oracles, and how the coordination between various features can be improved by a novel reward scheme. We present extensive experimental results to demonstrate the improved performance of the proposed methods over conventional feature selection approaches.

Learning pharmacogenomic multi-relations among diseases, genes and chemicals from content-rich biomedical and biological networks can provide important guidance for drug discovery, drug repositioning and disease treatment. Most of the existing methods focus on imputing missing values in the disease-gene, disease chemical and gene-chemical pairwise relations from the observed relations instead of being designed for learning high-order disease-gene-chemical multi-relations. To achieve the goal, we propose a general tensor-based optimization framework and a scalable Graph-Regularized Tensor Completion from Observed Pairwise Relations (GT-COPR) algorithm to infer the multi-relations among the entities across multiple networks in a low-rank tensor, based on manifold regularization with the graph Laplacian of a Cartesian, tensor or strong product of the networks, and consistencies between the collapsed tensors and the observed bipartite relations. Our theoretical analyses also prove the convergence and efficiency of GT-COPR. In the experiments, the tensor fiber-wise and slice-wise evaluations demonstrate the accuracy of GT-COPR for predicting the diseasegene-chemical associations across the large-scale protein-protein interactions network, chemical structural similarity network and phenotype-based human disease network; and the validation on Genomics of Drug Sensitivity in Cancer cell line dataset shows a potential clinical application of GT-COPR for learning diseasespecific chemical-gene interactions. Statistical enrichment analysis demonstrates that GT-COPR is also capable of producing both topologically and biologically relevant disease, gene and chemical components with high significance.