The front page of AI-driven biology
Thursday, October 15, 2026Thu, Oct 15
Updated 7:01 PM PT
Because laboratory tests rarely document proteins that fail to interact, computational biologists evaluate interaction models against randomly paired non-interacting proteins, an approach that systematically warps performance metrics.
University of Washington researchers have introduced RFdiffusion3, a generative model that uses free-atom diffusion to build custom protein and RNA structures around precise geometric constraints.
A multi-agent framework called ExpertTwins anchors language-model agents to inspectable evidence policies and cited publications rather than simple persona prompts.
Conditional generative models frequently hallucinate non-additive interactions when predicting how interventions behave in combination, according to an evaluation across drug dose-response, cellular perturbation, and molecular shape datasets.
Researchers have built a family of discrete diffusion models that construct phylogenetic trees from sequence data using top-down, middle-out, or bottom-up steps.
A computational model called SVC treats subcellular RNA transcript locations as image patches, using surrounding molecules to predict where individual genes localize within single cells.
Researchers introduced DNA-LLM2Vec, a staged conversion framework that equips directional, generative DNA language models with two-way sequence context.
Foundation models trained on general molecular chemistry successfully predict key properties of complex sugars, including immunogenicity and protein-glycan interactions.
Researchers have released MassNet, an AI-ready training dataset containing 1.5 billion tandem mass spectrometry spectra across 35 species, stored in a tensor format designed for deep learning.
Investigators developed Borzoi-SE, a genomic transformer framework that detects super-enhancers, large regulatory DNA clusters that drive cell identity, directly from long-range sequence across 90 human lineages.
A reinforcement learning framework called Multi-turn Feedback Guided optimization trains large language models to iteratively redesign molecules based on feedback from prior modification steps.
Researchers built CellChain, an evaluation benchmark spanning 590 biological cases and 2,360 matched conditions across gene dependency, cytokine signaling, and drug sensitivity.
Researchers demonstrated a black-box attack on the genomic language model Evo 2 7B that extracts memorized DNA sequences without access to model gradients, token probabilities, or training data.
Computational biologists created proteins designed from scratch (de novo proteins) that switch between two completely different three-dimensional folded states rather than settling into a single static structure.
Researchers mapped the binding profiles of 67 RNA-binding proteins in human cells using a ligation-free profiling method termed ecCLIP.
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.
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.
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.
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.
Researchers introduced USM-AA, a genomic foundation model that integrates reference genome coordinates directly into sequence representations.
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.
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.
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.
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.
Researchers built Prompteina, an all-atom generative model that creates protein structures from written prompts rather than specialized property encodings.
Researchers have adapted a video-segmentation foundation model, named ETSAM, to accurately map cell membranes within noisy 3D cryo-electron tomograms.
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.
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).
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.
Researchers combined two deep-learning approaches to overcome light-induced damage and slow acquisition in RESOLFT super-resolution microscopy.