AI tracks cell behavior in live video to pick out true blood stem cells
A deep-learning model trained on time-lapse video of single blood-forming stem cells predicted which cells could engraft long-term with 98.5% average precision.
Preprints
Recent original research from bioRxiv and arXiv. Preprints have not been peer reviewed.
Page 4
Structured agent plugin translates biological questions into graph database queries
Computational biologists created ddkg.skill, an agent module that compiles natural-language bioinformatics questions into executable Cypher code for the NIH Common Fund Data Distillery Knowledge Graph.
Combining structural models with sequence AI improves protein fitness predictions under distribution shifts
A machine-learning framework augments protein language model representations with structural priors and evolutionary co-variation features derived from an AlphaFold-style trunk.
Long-read assemblies reveal structural variation driving diversity in a corn fungus
Comparative analysis of four field strains of Colletotrichum graminicola against a reference assembly revealed extensive chromosomal rearrangements, duplications, and non-conserved minichromosomes.
Machine learning boosts directed evolution yields for engineered viral capsids 41-fold
Researchers combined machine learning with directed evolution to engineer adeno-associated virus (AAV) capsids optimized for exosomal encapsulation, a method used to protect gene therapy delivery vehicles from immune recognition.
Mathematical chemistry architecture guarantees valid molecules during reaction modeling
Researchers designed Neural Petri Flow, an architecture that embeds physical valence and atom-conservation laws as fixed mathematical layers while using neural networks only to predict reaction propensities.
Small shifts in AI protein embeddings let algorithms find fitter designs with fewer tests
Researchers showed that assay-specific fitness variations align within a compact, linear subspace of mutation representation changes produced by protein language models.
Generative model designs proteins, mRNA, and DNA simultaneously
Researchers built NuCaliby, a structure-conditioned generative model that designs protein amino-acid sequences together with their corresponding DNA and mRNA nucleotide sequences.
Hybrid 3D segmentation framework maps complex cell shapes in fluorescence microscopy
Nagini3D combines deep learning models with active surface algorithms to segment multiple objects directly across volumetric fluorescence microscopy images.
AI drug design agents match human proposals in historic pharma tests
Researchers built MAGI, an open modular agent designed to plan molecular optimization, launch generative tools, and interpret structure-activity relationships across drug discovery campaigns.
Language model cleans up generic answers in drug-repurposing knowledge graphs
Researchers built a scoring system called semantic precision score using the biomedical language model BioLORD to weed out overly broad answers, such as general terms like antibiotics or pharmaceutical preparations, from knowledge graphs querying potential treatments for 200 diseases.
Predicted structure of the CPLANE complex uncovers new partners in cilia formation
Researchers modeled the tertiary structure of the complete CPLANE complex and its largest subunit, JBTS17, using deep learning protein-folding tools.
AI feature spaces capture hidden cell death in Parkinson's disease models
Deep learning representations of induced pluripotent stem cell-derived neurons from three familial Parkinson's disease cases mirrored independent measurements of cell death, according to a bioRxiv preprint.
Protein language model predicts how molecular motors turn fuel into mechanical work
Researchers developed CatFunc, a lightweight protein language model pipeline that classifies how AAA proteins convert ATP chemical energy into mechanical force directly from sequence data.
Generative protein models help design targeted enzymes that strip phosphate tags from cancer and tau proteins
Researchers have developed peptide-retargeted phosphatases, synthetic fusion proteins that link computationally designed peptide binders to a serine/threonine phosphatase subunit to selectively remove phosphate groups from specific cellular targets.
Diffusion guidance helps build functional enzymes with multi-residue active sites
Bioengineers developed the Catalytic Guidance Engine, an iterative design framework that detunes geometric constraints during diffusion sampling to build complex catalytic sites into de novo proteins.
Discrete concept learning stabilizes spatial biology representations
Computational biologists introduced Kodiak, a discrete representation learning architecture that uses balanced codebook assignments to map complex scientific images into a reusable visual vocabulary.
Latent vocabularies help autoregressive models generate proteins faster
Researchers introduced two learned discrete latent vocabularies to train autoregressive transformers on protein sequence and structure, replacing conventional residue tokens and coordinates.
AI pipeline designs working enzymes from scratch for new-to-nature chemical reactions
Researchers introduced AlphaProtein Novo, a machine-learning pipeline that builds functional enzymes from scratch without starting from natural protein backbones.
Machine learning and simulations discover natural enzymes for green plastics
Researchers combined molecular simulations and machine learning across four screening rounds to prospect natural hydroxymethylfurfural oxidases for synthesizing the bioplastic precursor 2,5-furandicarboxylic acid (FDCA).
Microscope images fail to rescue spatial gene-expression foundation models from slide batch effects
Training spatial transcriptomic foundation models alongside matched tissue images does not consistently improve their ability to transfer cell annotations across donors, according to benchmarks across three models.
Tool turns single-cell data into executable gene-network models
A computational framework called scBOLT simplifies the process of turning single-cell RNA sequencing data across multiple experimental conditions into Boolean logic models of gene regulation.
Feedback-guided language model answers clinical trial questions with auditable evidence
Researchers introduced FD-SCoPE, an auditable language-model system designed to answer clinical questions using trial summary tables, including requests that require inferring unlisted drug targets or harmonized endpoints.
Generative flow model steers single-cell predictions across new perturbations
Researchers introduced scEGFlow, a framework combining conditional flow matching with energy guidance to model how individual cells shift in appearance and gene expression following biological perturbations.
Unsupervised graph method maps biomedical networks by tracking diffusion modes
Researchers built MOSAIC, an algorithm that converts complex biological graphs into compact coordinates by accumulating diffusion steps across the network rather than relying on standard eigenvalue sorting.
Transfer learning tailors genomic foundation model Borzoi to specific kidney cell types
Computational biologists created Kidzoi by fine-tuning the Borzoi sequence-to-function model on single-nucleus chromatin accessibility data from human kidney tissue.
Deep learning decodes DNA sequence rules linking nucleosome packing to cancer mutations
A deep-learning model named DeepND classified nucleosomal from linker DNA across the genome to reveal how spooling around histone proteins alters mutation and methylation rates.
Diffusion model sketches the shifting shapes of disordered proteins
A geometric diffusion model called Zephyr samples conformational ensembles of intrinsically disordered proteins directly from sequence, bypassing the need for slow physics simulations.
Lupus antibody atlas links disease flares to protein droplets and Epstein-Barr virus mimicry
Using phage display immunoprecipitation sequencing on 1,268 plasma samples from 695 lupus patients and 301 controls, researchers applied machine learning to define an expanded profile of linear autoantibody targets.
Organoid imaging and machine learning spot liver-toxic drugs
Researchers combined high-content imaging of patient-derived human liver organoids with machine learning to identify drugs that cause liver injury, classifying 54 known hepatotoxicants and 20 non-toxic controls.