Library

A comparison of five single-cell transcriptomics methods on human placenta reveals that sample processing drastically alters detected cell-type proportions, while in situ combinatorial indexing shows a strong, unexpected bias toward nuclear rather than cytoplasmic RNA.

CAMUS is a new scalable algorithm for estimating phylogenetic networks that can analyze datasets of up to 201 species in hours, overcoming the strict computational limits of previous methods.

AI protein structure prediction models like AlphaFold2 incorrectly predict that ionizable residues can be buried in nonpolar cores for mutated sequences, a physical impossibility that can be corrected using short molecular dynamics simulations.

The foundation model yHydra jointly embeds mass spectrometry spectra and peptide sequences into a shared space, enabling open and error-tolerant searches that increase peptide identification by up to 25% compared to existing methods.

GenVarLoader is a new dataloader that creates personalized genomic sequences on-the-fly, reading variants up to 26,000 times faster and reducing storage by 2,000-fold compared to existing methods.

The authors developed a multimodal contrastive learning model for Electronic Health Records that mathematically connects the loss function to the singular value decomposition of a pointwise mutual information matrix, enabling a novel privacy-preserving algorithm.

Network pharmacology provides a systems-level framework integrating multi-omics and tools like STRING and Cytoscape to overcome the limitations of single-target cancer therapies.

A new framework modeling morphological shape evolution as a diffusion process using Metropolis-Hastings MCMC improves ancestral shape estimation at the root and internal nodes of a phylogeny.

SciPhy is a new Bayesian phylogenetic tool that reconstructs highly accurate, time-scaled cell lineage trees and population growth dynamics from sequential CRISPR-based lineage tracing data.

Multiomic profiling of 86,689 single nuclei across 16 human tissues revealed that cell types delineated by DNA methylation differ extensively from those defined by 3D genome structure.

A deep learning model trained on hundreds of thousands of ECGs identified a novel electrocardiogram biomarker that predicts sudden cardiac death, flagging a high-risk patient group largely missed by current clinical standards.

A usability test demonstrated that high school students with no prior experience could successfully build and submit biological models using the MorpheusML ecosystem within a single workday.

The authors created RNASSTR, a massive dataset of over 4.7 million RNA sequence-structure pairs, and found that retraining existing deep learning models on this expanded data exposed severe generalization and computational limitations in current RNA folding architectures.

A single-cell transcriptomic atlas of over 6.3 million nuclei from 1,494 donors reveals shared cellular signatures across eight brain disorders and identifies specific neuronal, immune, and vascular changes linked to Alzheimer's disease severity.

The R package dreamlet uses precision-weighted linear mixed models on pseudobulk data to perform fast, memory-efficient differential expression analysis across large single-cell cohorts, demonstrated on 1.4 million nuclei from 299 Alzheimer's and control brains.

By integrating genomic, phenotypic, structural, and deep learning data, researchers mapped variants in the MCT8 thyroid hormone transporter to disease severity and developed a classifier that predicts variant pathogenicity (AUC 0.91) and severity (AUC 0.86) for over 8,000 variants.

By performing genome-scale perturb-seq on 22 million primary human CD4+ T cells, researchers mapped how perturbing every expressed gene alters transcriptomes in resting and stimulated states, identifying novel regulators of cytokine production and autoimmune disease risk.

General-purpose large language models can encode electronic health records converted into plain text to predict clinical outcomes just as accurately as specialized medical models, while avoiding the need for site-specific medical code vocabularies.

A computational tool called CURBD uses recurrent neural networks to accurately infer the direction and magnitude of communication currents between multiple brain regions from neural recording data.

By developing a new microfluidic assay called meSMiLE-seq, researchers identified DNA binding motifs for 99 previously uncharacterized human transcription factors and discovered that 14 of them specifically prefer binding to methylated DNA.

AlphaFold 3 struggles to accurately predict the structures of amyloid proteins, correctly identifying only 34% of known amyloids while falsely predicting amyloid structures for 54% of non-amyloid peptides, largely due to a bias toward globular proteins in its training data.

Single-nucleus RNA sequencing of the Alzheimer's disease middle temporal gyrus revealed sex-specific microglial signatures and identified MERTK genetic variation as a female-specific risk factor.

Using a novel machine learning framework called symclatron, researchers found that 15 to 23 percent of uncultivated bacteria and archaea across half of all known phyla likely engage in symbiotic lifestyles.

This paper proposes a theoretical three-stage model of lifespan aging driven by increasing biological entropy and introduces a conceptual Multiscale Entropic Aging Index (MEAI) to quantify systemic disorder across molecular, cellular, and tissue scales.

OrthoFinder v3 introduces a scalable, two-step phylogenetic algorithm that increases orthogroup inference accuracy by 5-7% and can process over 4,000 genomes, overcoming the quadratic time complexity bottleneck of previous methods.

The authors present a near-perfect telomere-to-telomere diploid genome benchmark for HG002 that adds 15.3 percent of previously unmapped genomic sequence and shows de novo assembly outperforms standard variant calling by an order of magnitude.

ESMDynamic is a deep learning model that predicts protein residue-residue contact dynamics directly from single sequences, matching or outperforming state-of-the-art ensemble methods like AlphaFlow while requiring orders-of-magnitude less computation.

The authors developed hoodscanR, a Bioconductor package that generates probabilistic, cell-level neighborhood profiles for spatial transcriptomics data to characterize mixed tissue environments.

Integrating multi-omics data with artificial intelligence accelerates precision drug discovery by shifting from single-target approaches to network-based models, significantly reducing development timelines and improving target identification.

Incorporating microbial genomic data into groundwater denitrification models via dynamic flux balance analysis and deep learning accurately predicts nitrate depletion and nitrite transients, outperforming conventional geochemistry-only models.

Toll-like receptors (TLRs) function as highly regulated, multi-layered information-processing hubs controlled by post-translational modifications, epigenetics, and metabolism, the dysregulation of which directly drives diseases like lupus and sepsis.

The Single-Cell Pediatric Cancer Atlas (ScPCA) Portal provides uniformly processed single-cell and single-nucleus RNA sequencing data for over 700 pediatric tumor samples across 55 cancer types.

After twenty years of development, the igraph 1.0 library has been released, providing a fast C-based core with Python, R, and Mathematica interfaces capable of analyzing networks with billions of edges.

A comprehensive benchmark of nanopore DNA methylation tools reveals that older models like Dorado v4r1 and RockFish are best for CpG sites, while newer Dorado v5 models excel at non-CpG and 6mA detection.

By representing phylogenetic trees in a tropical geometric space called "palm tree space," researchers achieved faster geodesic computations and better performance in statistical tasks like PCA compared to the standard BHV tree space.

By combining confirmed case counts and genetic sequencing data from US states between January 2021 and March 2022, researchers developed a model to jointly estimate the effective reproduction numbers and frequencies of co-circulating SARS-CoV-2 variants.

Co-Scientist, a multi-agent AI system built on Gemini 2.0, autonomously generates and refines novel scientific hypotheses that were successfully validated in wet-lab experiments for acute myeloid leukemia and liver fibrosis.

IQ-TREE 3 is a major update to a popular phylogenetic software tool, introducing advanced evolutionary models, sequence simulation, and tools for analyzing massive datasets like pandemic viral genomes.

BioMaster is a multi-agent framework utilizing Retrieval-Augmented Generation and role-based agents to automate complex bioinformatics workflows like RNA-seq and Hi-C processing more accurately than existing methods.

This review provides a comprehensive overview of large language models in bioinformatics, detailing their core components like tokenization and transformer architectures, and their applications across genomics, proteomics, and drug discovery.