The front page of AI-driven biology
Thursday, October 15, 2026Thu, Oct 15
Updated 9:31 PM PT
Recent original research from bioRxiv and arXiv. Preprints have not been peer reviewed.
Researchers built PeTriPPI2, a computational framework that predicts whether two proteins interact by pairing sequence embeddings from ESM-2 with structural features from PeTriBERT.
OmniVCBench introduces 6,077 question-and-answer pairs based on published figures to evaluate whether multimodal AI models can interpret wet-lab data and formulate biological hypotheses.
Researchers designed ZEST, a tokenization scheme that uses conserved regions from multiple sequence alignments to compress protein sequences into multi-residue tokens averaging four amino acids in length.
Researchers introduced CrystAF, an all-atom flow-matching model paired with the UMA interatomic potential, to evaluate the best stage to inject physics into generative models predicting molecular crystals.
Researchers developed PRIORITI, a computational framework that uses a domain-instructed large language model to synthesize genetic evidence into predictive representations for drug target and disease pairs.
Computational design yielded de novo mini-proteins that target Plasmodium falciparum reticulocyte-binding protein homolog 5 (RH5) to block its interaction with human basigin, an essential step in red blood cell invasion.
The Fusion2AI framework uses large language models to synthesize fragmented functional descriptions into standardized English annotations, Gene Ontology terms, and Enzyme Commission numbers for approximately 100,000 uncharacterized bacterial protein groups.
A Bayesian ordered logistic regression framework models minimum inhibitory concentrations as ordered intervals rather than simple resistant or sensitive categories in bacterial genome-wide association studies.
Deep learning models that predict gene expression levels directly from DNA sequences primarily capture whether genes are turned on or off rather than continuous, disease-relevant shifts in active genes.
PathoBERT combines a fine-tuned DNABERT encoder with multi-scale attention modules to predict whether short bacterial DNA fragments originate from pathogens.
Geneticists developed sum-h2, a fast statistical framework that scores the heritability of traits derived from deep learning without running full genome-wide association studies.
Bioinformaticians built ScProteoAgent, an artificial intelligence agent designed to translate natural-language prompts into reproducible analytical workflows for single-cell proteomics.
Researchers introduced MIMOSA, an inference-time framework that constrains generative protein diffusion models to mimic the interaction geometry of known binding partners.
Computational biologists at Tsinghua University developed MOD-scTWAS, a statistical framework that models co-expression modules to improve single-cell transcriptome-wide association studies.
Current quality-control heuristics for public genomic databases allow submitted ribosomal sequences to deviate by as much as 30% from known references, leaving public repositories vulnerable to contamination by synthetic or AI-generated data.
NeuronDiscover pairs prediction, intervention, and measurement design to untangle genuine biological responses from numerical modeling errors in computational digital twins of neuronal microenvironments.
Researchers introduced VenusREM-Harness, a plug-in method that improves mutation fitness predictions from frozen protein foundation models without retraining.
Researchers introduced M3OS, a framework that pairs large language models with Monte Carlo graph search to iteratively edit and improve drug-like small molecules.
KoopCell uses dynamical systems theory to infer cell development over time from unpaired single-cell RNA sequencing snapshots.
Researchers built an automated data-extraction pipeline to enrich 74,425 protein-ligand measurements in the public BindingDB database with structured experimental context pulled from original papers.
Researchers at the Oxford Protein Informatics Group derived a closed-form correction for a mathematical bias caused by pairing Dropout and LayerNorm layers, a common design in AlphaFold2-style architectures.
An evaluation of 14 frozen pathology foundation models found they fail to match a conventional convolutional neural network on neurodegenerative disease tissue, despite performing well on cancer biopsies.
Researchers built EpiZoo, a 2.6-billion-parameter mixture-of-experts model trained on 20.9 million single cells from multiple species to analyze chromatin accessibility, the physical opening of DNA for gene control.
A computational framework named BioPert uses measured gene expression changes in a single reference cell line, rather than chemical structure formulas, to predict how diverse cell lineages respond transcriptionally to drug exposure.
Researchers introduced GraPPI, a self-supervised deep learning model designed to represent protein-protein interactions by treating complexes as heterogeneous residual graphs.
A computational method named SpaReg aligns serial tissue slices to build three-dimensional tissue maps across histology, spatial proteomics, and spatial transcriptomics while correcting for physical tears and distortions.
TxCyto uses deep learning to estimate the signaling activity of 645 cytokines, which are immune signaling proteins, directly from standard RNA sequencing data rather than direct laboratory measurement.
Researchers built an algorithm called FOCAL, which pairs gene regulatory networks with interpretable machine learning to pick out the molecular subcircuits driving cell-state decisions.
Researchers built an XGBoost classifier trained on 42 genetic and biochemical features to detect effectors of the type VI secretion system, a molecular spear bacteria use to inject toxins into rival microbes or host cells.
Computational biologists built OmegaTherm, a sequence-based language model framework designed to jointly estimate shifts in folding free energy and melting temperature caused by single or multiple mutations.