Discover, rate, and discuss biology preprints.

The paper evaluates DNA foundation models for chimera detection in long-read metagenomic bins, finding that ensembling improves precision over single models.

This study evaluates class-balancing algorithms on metabolomics datasets from brain tumor patients and controls, finding that undersampling improves model performance while oversampling risks overfitting.

FRACTEL improves power and replication in CRISPR screens by analyzing region-level effects and informing experimental design.

PACMOS projects query cancer samples into pretrained MOFA latent factor spaces for classification.

PHI is a Galaxy-based workflow that automates prophage-host interaction analysis and generates standardized reports, detecting 41 prophages across 14 hosts in a mock community.

This study benchmarks AlphaFold and related methods for modeling antibody and TCR antigen recognition, showing improvements with AlphaFold3 and increased sampling but varying success across complex types.

OrthoGLMM tests associations between species traits and gene content changes while accounting for phylogeny, recovering signals for bacterial diazotrophy, plant nodulation, and marine mammals.

This study develops a transfer learning model to predict volatile-odorant receptor interactions in Diptera, revealing species-specific olfactory preferences and informing pest management strategies.

The paper evaluates zero-shot models' ability to predict epistatic effects on protein properties and finds they fail to predict strongly epistatic mutation combinations.

This paper introduces a method to inject distance information into AlphaFold-like structure predictors to improve structure prediction accuracy.

The paper presents a deep-learning pipeline for predicting natural product targets, identifying GPR35 as a potential target for ligiamycin A after correcting for bias.

The paper shows that accurate ΔTm prediction can be achieved using a sequence-only model without structural inputs, outperforming structure-informed methods.

RulePep is a neural-symbolic classifier for peptide-function prediction that combines ESM-2 representations with interpretable rules to achieve high performance on three peptide classification tasks.

The paper presents a high-quality acetylation dataset and shows that transfer learning can achieve good performance with minimal data for identifying PTMs.

Scop3P is updated to integrate phosphorylation sites with structural, functional, and mutational data, enhancing phosphosite interpretation and disease context.

RICO quantifies rDNA copy number and methylation using nanopore sequencing, showing accurate results in simulations and human/mouse samples, and revealing correlations between rDNA CN and methylation.

The paper introduces a standardized format for storing phased targeted sequencing data and provides tools to create and manipulate this format.

SpliSync improves RNA long-read alignment precision using a genomic language model, enhancing splice site correction but with limited impact on transcript reconstruction.

The paper predicts gene mutations in colon cancer using a gene-conditioned LSTM model with attention on inferred mutation paths from a directed co-occurrence asymmetry graph.

This paper develops an AI framework integrating single-cell RNA-seq and machine learning to analyze melanoma tumor microenvironment heterogeneity, identifying biomarkers and therapeutic targets.