benchmark study

Generative Models for Novel Enzymatic Reactions: How Far Do They Go?

This paper evaluates generative models for predicting novel enzymatic reactions, finding they can generate correct reactions for 25% of test cases and improve over existing methods.

Bioprocess Modeling Made Easy With Hybrax

Hybrax simplifies hybrid bioprocess modeling by providing a framework that automates model setup and data handling, enabling reproducible and reusable models.

Protify: A Benchmarking Platform Revealing Task Dependence and Homology Leakage in Protein Language Model Evaluation

Protify benchmarks protein language models across tasks and evaluation settings, revealing task dependence and homology leakage in model performance.

A comparative transcriptomic framework across kidney organoid, assembloid, and transplantation models

The paper compares transcriptomic profiles across kidney organoid, assembloid, and transplantation models to evaluate their maturity and suitability for research.

RPI-Zero: Zero-shot RNA-Protein Interaction Prediction

RPI-Zero predicts RNA-protein interactions without prior data by using a sequence-only model trained on a large benchmark of CLIP experiments.

memVar: an integrated resource for interpreting human membrane protein variation in molecular and disease contexts

memVar links human membrane protein variants to molecular contexts and disease, revealing enrichment of pathogenic variants in transmembrane regions and interaction sites.

Benchmarking Statistical Methods for Differential Composition Analysis in Population-Scale Single-Cell Data

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 enables benchmarking of cancer spatial transcriptomics methods across tumor evolutionary modes

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

Dated public evidence and known genes for retrospective evaluation of gene prioritization

The paper provides per-gene evidence tables from public resources for retrospective gene prioritization, including dates and leakage risk grades for four disease areas.

GPBench: A benchmark framework for genomic prediction

GPBench provides a standardized benchmark framework for genomic prediction, evaluating 21 models across 16 species and offering tools for analysis and comparison.

Decision benchmarking of cryo-EM re-refinement

The paper benchmarks cryo-EM re-refinement selection methods against no operation using a fixed cohort and scoring tools, finding increased disagreement with higher-resolution references.

Where single-cell foundation models help: a benchmark of annotation, perturbation and regulatory inference against simple baselines

This paper benchmarks single-cell foundation models against baselines for annotation, perturbation and regulatory inference using diverse datasets.

Autonomous generalizability auditing of scientific foundation models

This paper introduces AutoSPECTRA, a protocol for auditing generalizability of scientific foundation models across biological domains.

A Global Atlas of Immune Alleles Reveals Health-Relevant Variation Across Groups

The paper presents GAIA, a global atlas of immune alleles, showing that underrepresented groups have more diverse alleles needed for vaccine design than Northwest European groups.

Stable large-scale network architecture and structured individual variability in the awake mouse brain

The paper introduces the Awake Mouse Connectome Project (AMCP), a dataset of 212 awake mice, and presents functional network atlases and analysis tools for studying brain connectivity and variability.

Ark-ADMET: endpoint-specific selection of molecular representations and hyperparameters for gradient-boosted ADMET models under the Therapeutics Data Commons benchmark protocol

Ark-ADMET selects endpoint-specific molecular representations and hyperparameters for gradient-boosted ADMET models, achieving competitive performance without new architectures or shared representations.

BioResearchBench: 50 Long Horizon Biological Research Tasks

BioResearchBench evaluates AI models on 50 biology research tasks, revealing low success rates and highlighting challenges in data analysis, interpretation, and experimental design.

LabWeaver: reliable agentic data analysis grounded in laboratory knowled

LabWeaver improves agentic data analysis by organizing experimental context around laboratories, enhancing reliability in single-cell RNA-seq analysis of synovial sarcoma.

Synthetic MPRA perturbation supervision improves variant-effect prediction in sequence-to-function models

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

MillionTrees: A global benchmark for tree detection across geography, sensor, and annotation geometry

The paper introduces MillionTrees, a global benchmark dataset for airborne tree detection, showing that current models struggle to generalize across diverse data and that training on MillionTrees improves performance.

A multidimensional benchmark of confounder-aware differential abundance strategies for single- and multi-cohort microbiome studies

This paper benchmarks confounder-aware differential abundance methods for microbiome studies, finding that covariate adjustment improves false discovery control but reduces sensitivity, with different strategies performing better under various conditions.

Top2All: Benchmarking sparse-to-transcriptome reconstruction from high-expression profiles across sex- and sex-steroid-defined conditions

The paper benchmarks sparse-to-transcriptome reconstruction models using high-expression profiles from mouse adrenal cells under sex and sex-steroid conditions, finding that gene-wise metrics and strong baselines are critical for evaluation.

Leakage-Controlled Benchmarking of Variant Effect Predictors and Development of Gradient-Boosting Consensus Model

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

Quantum Barcoding Bridges High-Dimensional Immune Phenotyping with Signaling Responses across Species

The paper benchmarks Quantum Barcoding against mass cytometry in human and rhesus macaque blood, integrating data to reveal conserved and divergent signaling responses across species.

The Quagmire of Immunogenicity-Guided Antibody Design: Labels Limit Progress

The paper evaluates antibody design methods using immunogenicity labels and finds that format-based approaches outperform sequence-based predictors, but label inaccuracies limit progress.

RNA assignment and segmentation geometry in Visium HD profiles of fibrostenotic Crohn's disease

The paper evaluates segmentation methods for reconstructing cell-like profiles from spatial transcriptomics data but finds no reliable method under the given constraints.

Mapping bioimage-analysis software adoption in microscopy

The paper tracks the adoption of bioimage-analysis software across 1.7 million microscopy papers from 2010 to 2025, finding increased usage and concentration on general-purpose platforms like ImageJ/Fiji.

Negative sampling shapes benchmark discrimination and experimental retrieval in synthetic-lethality prediction

The paper evaluates how negative sampling affects synthetic-lethality prediction models, showing that benchmark performance does not reliably predict experimental retrieval success.

Toxicity Prediction Tools Do Not Generalize to Novel Chemistry

This study evaluates toxicity prediction tools on approved drugs, finding no method outperforms simple fingerprints, and highlighting the need for inputs beyond molecular structure.

A human induced pluripotent stem cell-based high-throughput screening platform for teratogenicity hazard identification

Four-model hiPSC screen in 384-well imaging reached 83% accuracy (88% sensitivity, 75% specificity) on 130 scored compounds from a 146-compound reference set, scalable to hundreds of compounds monthly.