AI infers tissue mechanical stress directly from cell shapes
StressNET, a graph neural network, estimates physical forces between neighboring cells directly from microscopy images without relying on pre-specified physical formulas.
Preprints
Recent original research from bioRxiv and arXiv. Preprints have not been peer reviewed.
Page 9
RNA language model decodes microRNA processing across 414 species
Researchers introduced miRstring, an RNA language model trained on 77,708 precursor sequences across 414 species to identify the boundaries of mature microRNAs (miRNAs), tiny RNA molecules that regulate gene expression.
Bayesian model maps cell types from sparse single-cell DNA methylation data
The Bayesian framework converts continuous methylation proportions into discrete states, factorizes the latent structure, and propagates measurement uncertainty into the final cell assignments.
AI pipeline audits biological interaction models against seven databases
University of Pittsburgh researchers developed BELL, a computational curation tool that combines rule-based retrieval from seven biological databases with large language models to verify mechanistic models.
Deep learning blends gene sequences and functional hierarchies to map cell networks
A computational biology team developed scMGFGRN, a framework that predicts gene regulatory networks by fusing single-cell RNA sequencing data with DNA sequence embeddings and Gene Ontology hierarchies.
Updated database adds 5.8 million predicted viral protein structures
The Steinegger lab released version 3 of the Big Fantastic Virus Database, expanding its collection of ColabFold-predicted viral protein structures to 5,776,417 models across UniProt sequences.
Protein language model unearths 14,300-year-old RNA viruses from frozen wolf remains
Scientists used the protein language model Lucaprot and AlphaFold2 structure modeling to detect two previously unknown RNA viruses in RNA sequencing data from a 14,300-year-old permafrost-preserved Pleistocene wolf.
Machine learning designs solid mRNA vaccines stable at body temperature
Researchers built an active-learning system called AGENT, combining Bayesian optimization with lab testing to formulate solid-state mRNA lipid nanoparticles in six iterative rounds.
AI tracks immune T cells to predict vaccine antibody responses
Researchers built a language model named DeepAir-Cov19 to spot SARS-CoV-2-targeting T-cell receptors, achieving an area under the curve of 0.94.
Projecting single-cell expression data onto a sphere yields interpretable developmental coordinates
Computational biologists developed SPHERE-PCA, a method that maps the first three principal components of single-cell RNA sequencing data onto a three-dimensional unit sphere to create an interpretable coordinate framework.
Adversarial machine learning predicts peptide functions without confirmed negative data
Researchers built Pep-PU-GAN, a machine learning model that pairs graph neural networks with positive-unlabeled adversarial learning to classify bioactive peptides when verified non-functional examples are scarce.
Spatial maps of 112 million melanoma cells pinpoint immune hubs that predict immunotherapy response
A spatial multi-omics study tracking over 112 million cells across three melanoma neoadjuvant immunotherapy regimens revealed that mature tertiary lymphoid structures with germinal centers, sustained by specialized fibroblast scaffolds, distinguish responders from non-responders.
Generative model translates protein sequence into microscopy images
Researchers built CELL-FM, a bidirectional generative AI framework that predicts fluorescence microscopy images directly from protein sequences and cellular context, and can also work in reverse.
Sparse machine learning maps cord blood omics in infant lung disease
Researchers applied a sparse machine learning workflow using Stabl and LASSO regression to analyze 45,000 cord blood features in a cohort of 217 infants at risk for bronchopulmonary dysplasia, a chronic lung condition affecting premature babies.
Language model decodes complex fat molecules in mass spectrometry data
Researchers built OpenLipid, a workflow that uses large language models to identify target lipids from data-independent acquisition mass spectrometry.
Machine learning trained on tumor gene loss spots potent cancer neoantigens
Researchers built the Neoantigen Elimination Model (NEMo), a two-compartment machine learning tool that predicts which mutant tumor targets trigger immune attack by tracking which cancer subclones disappear during immunotherapy.
AI turns text descriptions of genes into an interpretable coordinate system for single-cell data
A computational method named MESIC uses a biomedical language model to translate written gene annotations into low-dimensional semantic coordinates that give biological meaning directly to single-cell data.
Generative model designs synthesizable drug candidates validated across five protein targets
Researchers introduced the Large Drug Discovery Model, a 3D generative framework that designs drug candidates to satisfy fine-grained structural goals while ensuring they can be readily made in a laboratory.
Generative search guides protein design to build binders for carbohydrates and polar surfaces
Researchers introduced latent generative search, a framework that guides the Proteina-Complexa model at inference time to generate protein sequences and 3D shapes simultaneously, eliminating the need for a separate inverse-folding step.
Adding physical rules brings three-dimensional drug-binding models up to standard ligand baselines
A computational study evaluated whether adding structural and physical rules can help three-dimensional machine learning models match conventional two-dimensional ligand-only models in predicting small-molecule binding affinity.
Fine-tuned neural networks predict molecular toxicity in data-scarce chemical classes
Researchers fine-tuned a general baseline molecular property model on small, specialized datasets to predict chemical toxicity in classes with limited historical testing.
Clustering tool for single-cell genome architecture separates true gaps from missing data
A computational method called ssJSD groups single-cell chromatin contact maps by distinguishing biologically meaningful uncontacted DNA regions from technical dropouts caused by shallow sequencing.
Autonomous laboratory robot tests physical inputs on living cells and embryos
Researchers built MOMbot, an automated hardware and software platform that delivers chemical, electrical, mechanical, and thermal cues to biological specimens while an active-learning algorithm decides the next test.
Blood-protein survival model predicts 10-year risk across hundreds of diseases simultaneously
Scientists developed MMAD-Risk, a multivariate survival model designed to predict long-term risk across 271 diseases simultaneously from blood plasma proteomic profiles.
Physics-based machine learning predicts enzyme reaction rates and selectivity
Researchers combined machine-learning models of enzyme electrostatics with molecular mechanics and microkinetic modeling to simulate multistep catalytic reactions without relying on empirical approximations.
Genomic language model designs engineered enzyme assembly lines
Bioengineers coupled the genomic foundation model gLM2 with discrete diffusion to guide the generative design of complex, multi-domain biosynthetic gene clusters.
Matrix factorization reveals 200 distinct gene programs across mouse T cells
The immgenT consortium developed immgenT-GP, a computational framework that deconvolves roughly 800,000 mouse T cells into 200 distinct, reproducible gene expression programs.
AI coding assistants struggle to build correct sequencing pipelines without human oversight
A benchmark evaluated three automated coding agents, Biomni, Claude Code, and Codex, on writing a complex Nextflow pipeline for processing epigenomic sequencing data from start to finish.
Designed mini-proteins block Nipah and Hendra viruses in lab tests
Computational biologists generated thermostable, de novo mini-proteins that bind Nipah and Hendra virus surface glycoproteins with subnanomolar affinity.
ModelSEED metabolic database adds reliability tiers and language-model direction calls
An update to the open-source ModelSEED Biochemistry Database expands the resource to approximately 46,000 compounds and 56,000 mass- and charge-balanced reactions for whole-cell metabolic modeling.