The front page of AI-driven biology
Thursday, October 15, 2026Thu, Oct 15
Updated 6:32 PM PT
Researchers adapted the DINOv3 vision transformer to classify five common bacterial pathogens from phase-contrast time-lapse imaging of cells grown in microfluidic traps without fluorescent dyes.
Researchers compared the specialized splicing model OpenSpliceAI against the genomic foundation model AlphaGenome to predict rare, disease-linked RNA splicing errors called cryptic exons caused by TDP-43 dysfunction.
Researchers combined a phenotypic screen of around 1,200 approved drugs with Boltz-2 structural modeling to identify inhibitors of NALCN, a sodium leak channel mutated in the ultra-rare neurodevelopmental condition CLIFAHDD.
Researchers introduced MD-LLM-2, a transferable language model that generates protein motion trajectories conditioned on sequence, temperature, and structural history.
Alphabet spin-out Isomorphic Labs is in preliminary discussions to raise external capital at a valuation of at least $40 billion, according to reporting from Bloomberg.
Researchers developed PK-MUSE, a computational model that marries traditional two-compartment physiological equations with neural networks to predict human drug concentration over time directly from chemical structure.
Computational biologists evaluating metamorphic proteins, which switch between distinct three-dimensional folds, found that protein language models rely heavily on sequence memorization and fail to generalize to novel shape-shifting regions.
Researchers built REFinder, a pipeline that uses a DNA language model called REBEAN to label raw sequencing reads by enzyme class and assemble only sequences with putative catalytic signatures.
The Chan Zuckerberg Biohub has unveiled an expanded $1.8 billion Virtual Biology Initiative in collaboration with Meta, Google DeepMind, Isomorphic Labs, the Department of Energy, and the National Institutes of Health.
Computational biologists introduced Kodiak, a discrete representation learning architecture that uses balanced codebook assignments to map complex scientific images into a reusable visual vocabulary.
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.
Computational biologists created Kidzoi by fine-tuning the Borzoi sequence-to-function model on single-nucleus chromatin accessibility data from human kidney tissue.
A computational benchmark evaluated ten large language models on linking 5,861 compounds from the ChEMBL database to 14 adverse outcome pathways spanning five neurodevelopmental toxicities.
Researchers introduced AuditPPI, an interpretability framework that converts sparse-autoencoder representations from frozen protein language models into auditable pair fingerprints to evaluate how models predict protein-protein binding.
Researchers have built RipplePLM, a computational model that generates natural language descriptions of how single-amino-acid changes disrupt protein behavior.
An analysis of approximately 2 million PubChem bioassays found that more than 99.9% lack standardized ontology tags for assay formats and detection technologies, creating major hurdles for training molecular property models.
A systematic benchmark evaluated 4-bit quantization and low-rank adapter fine-tuning, known as QLoRA, across seven protein language models including ESM-2, ProtT5, and Profluent-E1.
A team retrained a protein language model classifier on experimentally mapped proteomes from eight diverse species to predict the mitochondrial proteins of roughly 200 eukaryotes.
Researchers built AREION, a transformer-based brain foundation model that learns functional magnetic resonance imaging representations regardless of which brain atlas was used to define anatomical regions.
Researchers built BioEvidence, an interface that feeds structural and evolutionary context from standard biology tools into frozen large language models to predict how protein mutations affect function.
Researchers built CellMSA, a single-cell foundation model that borrows the concept of multiple sequence alignments in proteins to capture gene relationships across different cell types and batches.
Researchers built Mendel, an AI foundation model trained from scratch on human whole-genome sequences using a training objective focused on differences between individual genomes rather than invariant sequence.
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 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.
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.
PathoBERT combines a fine-tuned DNABERT encoder with multi-scale attention modules to predict whether short bacterial DNA fragments originate from pathogens.
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.
Researchers introduced VenusREM-Harness, a plug-in method that improves mutation fitness predictions from frozen protein foundation models without retraining.
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.