transformer / foundation model

This study evaluates genomic foundation models for viral sequence classification, finding that NTv3-650M achieves high accuracy comparable to BLAST on standard-length sequences.

This paper develops a deep learning model to identify drug modes of action from high-throughput images of chemically and genetically perturbed Corynebacteria.

Minerva extracts RNA base-pairing, protein-contact and repeat maps from a genome language model without alignments, flagging unannotated structured loci and a phage ncRNA array reverse-transcribed into short cDNAs.

ASOFormer predicts ASO efficacy using sequence, structure, and modification features, outperforming existing methods in accuracy and candidate prioritization.

Continued pretraining of an RNA language model on picornaviral 5′ UTRs alone predicts IRES base pairs with median precision 0.80 versus 0.47 for Infernal, and reveals a Type II subclass.

Bidirectional Mamba2 plant genomic language models (318M to 3.2B parameters) trained on 320 land-plant genomes top a 22-task frozen-model aggregate; a small variant reaches 131,072 bp context.

DeepSCENIC uses transfer learning from sequence-to-function models to infer causal gene regulatory networks from single-cell multiome data.

The paper determines the structure of norovirus NS3 and proposes it functions as an RNA translocase.

This paper uses a transformer-based model to show that macaque V1 neurons use sparse, stimulus-dependent readout for orientation information.

This study benchmarks biomedical language models for concept normalization from real-world text, finding Qwen 2 Instruct improves accuracy when combined with SapBERT.

SAP provides a transferable, training-free framework for automated segmentation of plant time-series imagery across species and conditions.

This paper develops a transformer-based classifier to detect artificially edited bacterial genomes by analyzing gene family contexts, achieving accurate identification without relying on specific marker genes.

Reward-guided inference-time search over the Proteína-Complexa sequence-structure codesign model yields validated binders across 127 targets, kinases and a free carbohydrate, with codesigned sequences outperforming ProteinMPNN redesign.

LLMsFold designs and validates new small molecules targeting pathogenic proteins using LLMs and biophysical simulations, showing strong in silico binding and drug-like profiles.

This study evaluates two time-series foundation models for cognitive workload classification from eye-tracking data, showing better generalization to unseen individuals than task-specific models.

The paper introduces RSAUNet, a deep learning model for prostate cancer segmentation in MRI scans, achieving high Dice and Jaccard scores.

Drone imagery from 356 flights in 19 Genomes to Fields maize environments; vegetation-index functional principal components complement genomic kernels, predicting yield of held-out hybrids at r up to 0.501.

BoneGraph combines a SPECTER2-indexed bone corpus, knowledge graph, deterministic physics checks, and user-taught rules in a local five-tab system; oracle passages raise accuracy 42% to 78%.

This study uses time-series foundation models to improve forecasts of lake phytoplankton dynamics, finding that Chronos-T5 outperforms other models in accuracy and stability.

The paper identifies maternal immune signatures in repeated pregnancy loss using single-cell RNA sequencing and machine learning, linking them to potential drug repurposing candidates.

This study shows that Brf1-mediated RNA polymerase III activity restricts murine gammaherpesvirus spread by enhancing host antiviral responses during low MOI infections.

Molecular oncology papers that textually resemble retracted paper mill papers receive 50 to 100% more citations yet fewer readers, cite one another, and measurably inflate some journals' citation rates.

On 495 human Patch-seq neurons, cell type means predicted electrophysiology better than neural networks on scGPT embeddings or gene panels; fusing scGPT embeddings with cell type labels predicted best.

This paper introduces TomoFLIM, a high-speed multiphoton FLIM technique that enables rapid, accurate imaging of dynamic biological processes with compression ratios over 90%.

Seven high-speed cameras, a hybrid 2D/3D network and body-model inverse kinematics track 50 keypoints on freely moving flies; height and speed peak together at all running speeds.

The paper proposes a computational framework where cortico-subcortical circuits implement multi-head self-attention, aligning with human brain recordings during speech perception.

The paper introduces Xpressor, a framework for cross-scale learning in biology, improving cell-type prediction and embedding quality.

REN-former identifies gene regulators of kidney disease transitions by combining single-cell modeling and human genetics.

TransBind2 improves TF-DNA binding prediction by integrating TF structure, chromatin context, and bidirectional cross-attention, achieving high accuracy across human ChIP-seq data and cross-species predictions.

This paper redesigns PET hydrolase PHL7 using MPNN models to enhance its catalytic activity at lower temperatures, achieving efficiency comparable to its activity at higher temperatures.