benchmarking

The paper derives a scaling law for biological learning by linking inductive bias to learning speed and validates it using human and mouse datasets.

This study shows that excluding low-frequency variants and indels in fine-mapping leads to false discoveries and introduces SuSiE-OUT to address this issue.

This paper introduces SHARP, a method using cosmological web analysis to characterize brain ageing and cognitive impairment by analyzing MRI data from CamCAN and OASIS-3.

This study evaluates AlphaGenome's ability to predict gene expression changes after promoter perturbations, finding discrepancies between CAGE and RNA-sequencing predictions.

BioMeme predicts cell responses to new perturbations by recalling measured experiments, outperforming neural models in accuracy and speed.

The paper introduces a hub labeling method to reduce the space required for distance indexing in large pangenome graphs, improving query support and reducing storage needs.

The paper shows that plate design confounds donor variation in single-cell perturbation screens and demonstrates how improved analysis methods can change interpretations of recent studies.

This study demonstrates high reproducibility in a DTC microbiome testing workflow using NIST fecal reference material across multiple processing steps and real-world conditions.

This study benchmarks 12 methods for differential composition analysis in scRNA-seq data, finding LOCOM performs best in controlling false discovery rate and maintaining power.

OncoSpot benchmarks cancer spatial transcriptomics methods across tumor evolutionary modes by simulating realistic cancer data with known subclonal and copy-number features.

The paper assesses the transportability of mouse gene knockout models to human complex phenotypes using GPAT framework, finding evidence for 12 phenotypes but limited for others.

This study compares age-specific pediatric EEG templates to adult templates and finds they better match individual MRI-based biomarkers in children up to 12 years old.

This paper introduces an end-to-end in silico framework to validate and test clinical genome sequencing pipelines from simulated reads to reportable findings using a fixed truth set.

SPaRTA enables non-experts to query spatial-omics data via natural language, achieving 63-77% accuracy with fine-tuning on cancer tissue datasets.

This paper evaluates how site heterogeneity affects brain-phenotype predictions in multi-site data, finding that larger sample sizes improve accuracy more than multi-site composition.

Rho is a collaborative analysis environment that organizes materials by scientific purpose to support human-AI judgment in bioinformatics.

The paper evaluates pesticide survival models by analyzing lesion persistence, showing that post-exposure death rates better reflect hazard structure than peak-based assumptions.

KDRD estimates replicability of associations using summary statistics from a single study, showing moderate to high replicability for 33 of 45 brain-related traits in proteomics analyses.

This paper introduces vTRM, a framework to characterize WMH spatial patterns across age and sex, revealing distinct associations with aging, vascular, and neurodegenerative processes.

The paper shows cancer classifiers can achieve perfect accuracy by learning data source shortcuts rather than biological features, and introduces a method to detect and mitigate this issue.

The paper introduces a longitudinal neuroimaging dataset of Alzheimer's disease pathology, including multimodal data from patients and controls, to identify and validate novel biomarkers.

This paper introduces ORCA, a platform that uses oncogenic RNA sensing to target and ablate cancer cells, demonstrating efficacy in colorectal cancer models.

This study compares and validates two BIDS Apps, PETPrep and PICNIC, against in-house software for preprocessing UCB-J PET data, finding high consistency but region-dependent bias.

This paper improves variant-effect prediction in sequence-to-function models using synthetic MPRA data generated from distilled model predictions.

The paper introduces eQIF, an unsupervised machine-learning method that integrates Enformer predictions to prioritize gene-regulatory variants, showing strong performance in predicting eQTLs and GWAS variants.

This paper evaluates the accuracy of structural MRI in classifying autism using leakage-free, site-aware, and externally validated machine learning methods across two ABIDE cohorts, finding modest performance.

SpaceBlocks is a workflow for reproducible, expert-supervised spatial transcriptomics analysis across platforms, identifying new cell types and differentially expressed genes in multiple datasets.

The paper benchmarks variant effect predictors using a leakage-controlled dataset and develops a gradient-boosting consensus model that outperforms existing tools.

The paper shows that the human gut microbiota recovers to an individual-specific state after antibiotics, not to a generic healthy state.