Large Language Models (LLMs) offer substantial promise for clinical natural language processing (NLP); however, a lack of standardized benchmarking methodologies limits their objective evaluation and practical translation. To address this gap, we introduce ClinBench, an open-source, multi-model, multi-domain benchmarking framework. ClinBench is designed for the rigorous evaluation of LLMs on important structured information extraction tasks (e.g., tumor staging, histologic diagnoses, atrial fibrillation, and social determinants of health) from unstructured clinical notes. The framework standardizes the evaluation pipeline by: (i) operating on consistently structured input datasets; (ii) employing dynamic, YAML-based prompting for uniform task definition; and (iii) enforcing output validation via JSON schemas, supporting robust comparison across diverse LLM architectures. We demonstrate ClinBench through a large-scale study of 11 prominent LLMs (e.g., GPT-4o series, LLaMA3 variants, Mixtral) across three clinical domains using configurations of public datasets (TCGA for lung cancer, MIMIC-IV-ECG for atrial fibrillation, and MIMIC notes for SDOH). Our results reveal significant performance-efficiency trade-offs. For example, when averaged across the four benchmarked clinical extraction tasks, GPT-3.5-turbo achieved a mean F1 score of 0.83 with a mean runtime of 16.8 minutes. In comparison, LLaMA3.1-70b obtained a similar mean F1 of 0.82 but required a substantially longer mean runtime of 42.7 minutes. GPT-4o-mini also presented a favorable balance with a mean F1 of 0.81 and a mean runtime of 13.4 minutes. ClinBench provides a unified, extensible framework and empirical insights for reproducible, fair LLM benchmarking in clinical NLP. By enabling transparent and standardized evaluation, this work advances data-centric AI research, informs model selection based on performance, cost, and clinical priorities, and supports the effective integration of LLMs into healthcare. The framework and evaluation code are publicly available at https://github.com/ismaelvillanuevamiranda/ClinBench/.

Biological network-based strategies are useful in prioritizing genes associated with diseases. Several comprehensive human gene networks such as STRING, GIANT and HumanNet were developed and used in network-assisted algorithms to identify disease-associated genes. However, none of these networks are disease-specific and may not accurately reflect gene interactions for a specific disease. Aiming to improve disease gene prioritization using networks, we propose a Disease-Specific Network Enhancement Prioritization (DiSNEP) framework. DiSNEP first enhances a comprehensive gene network specifically for a disease through a diffusion process on a gene-gene similarity matrix derived from disease omics data. The enhanced disease-specific gene network thus better reflects true gene interactions for the disease and may improve prioritizing disease-associated genes subsequently. In simulations, DiSNEP that uses an enhanced disease-specific network prioritizes more true signal genes than comparison methods using a general gene network or without prioritization. Applications to prioritize cancer-associated gene expression and DNA methylation signal genes for five cancer types from The Cancer Genome Atlas (TCGA) project suggest that more prioritized candidate genes by DiSNEP are cancer-related according to the DisGeNET database than those prioritized by the comparison methods, consistently across all five cancer types considered, and for both gene expression and DNA methylation signal genes.

MOTIVATION: MicroRNAs (miRNAs) are small non-coding RNAs that have been successfully identified to be differentially expressed in various cancers. However, some miRNAs were reported to be up-regulated in one subtype of a cancer but down-regulated in another, making overall associations between these miRNAs and the heterogeneous cancer non-linear. These non-linearly associated miRNAs, if identified, are thus informative for cancer subtyping. RESULTS: Here, we propose mirPLS, a Partial Linear Structure identifier for miRNA data that simultaneously identifies miRNAs of linear or non-linear associations with cancer status when non-linearly associated miRNAs can then be used for subsequent cancer subtyping. Simulation studies showed that mirPLS can identify both non-linearly and linearly outcome-associated miRNAs more accurately than the comparison methods. Using the identified non-linearly associated miRNAs much improves the cancer subtyping accuracy. Applications to miRNA data of three different cancer types suggest that the cancer subtypes defined by the non-linearly associated miRNAs identified by mirPLS are consistently more predictive of patient survival and more biological meaningful. AVAILABILITY AND IMPLEMENTATION: The R package mirPLS is available for downloading from https://github.com/pfruan/mirPLS. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

MOTIVATION: Recent technology developments have made it possible to generate various kinds of omics data, which provides opportunities to better solve problems such as disease subtyping or disease mapping using more comprehensive omics data jointly. Among many developed data-integration methods, the similarity network fusion (SNF) method has shown a great potential to identify new disease subtypes through separating similar subjects using multi-omics data. SNF effectively fuses similarity networks with pairwise patient similarity measures from different types of omics data into one fused network using both shared and complementary information across multiple types of omics data. RESULTS: In this article, we proposed an association-signal-annotation boosted similarity network fusion (ab-SNF) method, adding feature-level association signal annotations as weights aiming to up-weight signal features and down-weight noise features when constructing subject similarity networks to boost the performance in disease subtyping. In various simulation studies, the proposed ab-SNF outperforms the original SNF approach without weights. Most importantly, the improvement in the subtyping performance due to association-signal-annotation weights is amplified in the integration process. Applications to somatic mutation data, DNA methylation data and gene expression data of three cancer types from The Cancer Genome Atlas project suggest that the proposed ab-SNF method consistently identifies new subtypes in each cancer that more accurately predict patient survival and are more biologically meaningful. AVAILABILITY AND IMPLEMENTATION: The R package abSNF is freely available for downloading from https://github.com/pfruan/abSNF. SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.