RNA-protein binding atlas trains models to read structural genetic rules
Researchers mapped the binding profiles of 67 RNA-binding proteins in human cells using a ligation-free profiling method termed ecCLIP.
Preprints
Recent original research from bioRxiv and arXiv. Preprints have not been peer reviewed.
Page 2
Generative models design candidate molecules to reverse nerve-agent poisoning
Researchers benchmarked generative design pipelines to create non-oxime reactivators of human acetylcholinesterase, an essential enzyme disabled by organophosphate nerve agents and pesticides.
New computational tool extracts mitochondrial mutations from standard single-cell RNA sequencing
A computational tool called scMOCHA recovers mitochondrial DNA variants and heteroplasmy directly from standard single-cell RNA sequencing datasets.
Foundation model reads spontaneous electrical activity in brain organoids
Researchers trained an electrical foundation model named PULSE on more than 900 hours of microelectrode array recordings from lab-grown neural tissue and brain organoids.
AI groups genes into biological pathways to compress single-cell foundation models
Researchers built BioMod, a 50-million-parameter single-cell model trained on roughly 73 million human cells using 660 curated pathways and gene programs rather than thousands of individual genes as its basic input tokens.
Language models fail to beat standard algorithms at designing ionizable lipids
Large language models used to suggest new ionizable lipid molecules showed no meaningful edge over standard genetic algorithms or random search under matching experimental budgets.
Dynamic metabolic modeling predicts fecal transplant engraftment and resistance to C. difficile
Researchers built muODE, an automated computational framework that converts reconstructed bacterial genomes into dynamic metabolic simulations integrated with ecological factors such as pH, antibiotic levels, and bile-acid transformations.
GPU tensor reformulations accelerate structural docking analysis in HADDOCK3
Developers restructured the pairwise structural comparison algorithms in the macromolecular docking suite HADDOCK3 into continuous GPU tensor operations.
Single-cell foundation models beat simple baselines only when labeled cells are scarce
Benchmarking cell representations from four single-cell foundation models across 868 disease-prediction tasks and 65 cell-type datasets showed that standard log-normalized gene counts match or outperform the best model, TranscriptFormer-Sapiens, when full training data is available.
Genomic language model weaves reference coordinates into DNA sequence modeling
Researchers introduced USM-AA, a genomic foundation model that integrates reference genome coordinates directly into sequence representations.
Machine learning picks better target sites for programmable RNA sensors
Computational biologists developed DeSTAR, a machine-learning model that ranks target sites on RNA transcripts to improve the performance of toehold-switch RNA sensors.
Statistical model maps full amino acid flexibility beyond classical Ramachandran plots
Researchers modeled the complete torsional angles of all 20 amino acids and 400 amino-acid pairs across up to 14 dimensions, conditioning the continuous distributions on secondary structure elements.
Deep learning toolkit decodes raw nanopore signals from native tRNA
Computational biologists built tRNAZAP, a transformer-based toolkit that directly analyzes raw ionic current signals from nanopore sequencers to read short, heavily modified transfer RNA molecules.
General AI models reason through protein mutations without specialist tools
Researchers benchmarked whether general large language models can predict mutation effects directly from raw three-dimensional atomic coordinates without relying on dedicated molecular energy calculators.
Single-cell foundation models often perform better using middle layers than final outputs
Benchmarking four major single-cell foundation models across cell-type and perturbation tasks revealed that intermediate layers frequently match or outperform the final layer representations standardly used in downstream biology.
AI designs protein structures directly from text descriptions
Researchers built Prompteina, an all-atom generative model that creates protein structures from written prompts rather than specialized property encodings.
Single-cell pipeline infers gene regulatory networks and cell types simultaneously
Computational biologists introduced NetMap, a software pipeline that jointly infers single-cell gene regulatory networks and clusters cell populations without needing pre-assigned cell annotations.
Protein language models map membrane interaction surfaces from sequence alone
Researchers developed CHORUS, a deep-learning method that identifies binding interfaces on membrane proteins directly from raw amino acid sequences without requiring three-dimensional structures or evolutionary alignments.
Common AI confidence scores fail to predict whether designed binders hit their targets
Computational binder screening frequently filters candidates by predicted interface confidence, but a benchmark of 196 nanobody loop designs targeting the cancer protein survivin found no correlation between confidence metrics and actual epitope contact (rho = +0.100).
Machine learning scores cell maturity from DNA methylation marks
EpiDiff uses machine learning on DNA methylation data to produce a continuous numerical score of cellular differentiation across single-cell, sorted, and bulk tissue samples.
Multi-stage reasoning model designs catalytic enzymes from chemical reactions
A team led by Scinetics introduced chain-of-structures, a framework that couples language reasoning with structural generative models to design enzymes directly from reaction specifications.
Generative model designs all-atom peptides directly inside binding pockets
Researchers at Tsinghua University built PeTFlow, a generative flow-matching model that designs binding peptides directly within three-dimensional protein pockets.
Tool shrinks microbiome protein databases using direct peptide predictions
FastaLake uses AlphaNovo de novo peptide predictions to retrieve complete microbial proteins from large catalogs, generating sample-specific search spaces that match or outperform traditional metagenomic databases.
Autonomous AI agents mine metagenomic data to discover new viral enzyme families
Researchers deployed autonomous instances of Claude Code to search 1.9 billion protein clusters for reverse transcriptases, identifying a new class of giant-virus enzymes dubbed array-associated RTs.
Protein design model combines ligand contacts and evolution to boost enzyme activity
A graph neural network called zPolaron designs functional proteins by modeling interactions between protein backbones and bound ligands alongside evolutionary patterns from ESM-2.
Neural network links cancer mutations directly to shifts in DNA chemical marks
A machine learning tool called unduliner uses long-read sequencing data to connect specific somatic mutations with local changes in DNA methylation patterns.
Benchmark overhaul shows simple algorithms beat deep learning in picking gene perturbation targets
A new evaluation framework named Cause-excluded Effect Evaluation exposes that previous single-cell target discovery benchmarks were inflated by data leakage and by leaking the target gene's own expression.
Updated AI model predicts full RNA transcripts and pinpoints disease-causing splicing errors
Researchers have introduced SpliceAI2, a deep learning model trained across multiple mammalian species that reconstructs full-length transcripts directly from DNA sequences.
Massive cellular screen maps half a million mutations across cancer signaling pathway
Scientists measured the functional impact of 186,676 protein variants across 17 enzymes in the RAS-MAPK signaling cascade, generating 503,099 total cellular readouts.
AI transfer model maps noncoding variant effects across developing human brain regions
Seal, a new transfer-learning framework, predicts spatiotemporal gene regulation and variant impacts across 26 brain regions, 30 cell types, and seven human developmental stages.