transformer / foundation model

This paper shows that biological concepts can be represented as directions in foundation model embeddings, enabling interpretable and reusable analysis of diverse biological properties.

FOCUS improves spatial transcriptomics by coordinating molecular and spatial fidelity across multiple tasks and datasets, revealing clearer biological patterns in immunology, development, and oncology.

This paper introduces NanoPCT, a framework that predicts protein corona formation by learning universal patterns across proteins and enables zero-shot prediction for unseen proteins.

Linear NeuroPaint evaluates a linear model for cross-session neural alignment and inpainting, finding it less accurate than nonlinear models but computationally efficient.

UniMedSeg improves multi-modal medical image segmentation by integrating uncertainty-guided encoding and dual-prompt decoding for better domain generalization and spatial risk estimation.

This paper develops a vision transformer model for rapid bacterial species identification from time-lapse images without fluorescent labels, achieving high accuracy on clinical isolates.

EventHorizon is a foundation model for clinical flow cytometry that uses self-supervised learning to generate embeddings for diagnostic use, achieving high accuracy in identifying genetic abnormalities.

The paper develops a vision normalizing flow model to detect banana diseases from in-field images with high accuracy, using pre-trained embeddings and probabilistic density estimation.

The paper predicts viral host range using a combination of experimental screening and deep learning to identify potential animal reservoirs for reverse zoonotic transmission.

The paper introduces Cerberus, an improved regulatory sequence model using bidirectional state space blocks, achieving higher accuracy and efficiency than existing models like Borzoi.

VesiScope detects and analyzes GUVs from phase contrast images using a YOLOv11 model, enabling accurate and efficient label-free quantification.

GRASP is a graph Transformer that learns molecular structure from ZINC20 data and adapts to ChEMBL bioactivity, achieving better performance than existing models on ADMET prediction tasks.

STEM-LM uses masked language modeling to predict multi-species distributions by treating species presence as tokens in a spatio-temporal 'sentence', achieving high discrimination on ecological datasets.

Predicting differentially expressed genes and expression shifts per condition from database embeddings plus per-gene evidence, PIE reaches 1.2–3.2 times the best baseline's AUPRC on unseen contexts, perturbations and datasets.

This study evaluates zero shot, one shot, and few shot prompting regimes for transformer-based synthetic health tabular data generation using small models, finding zero shot improves fidelity but reduces privacy.

The paper uses AlphaFold2 to model class B β-lactamases without experimental structures, tests the BBL numbering scheme's validity, and identifies structural exceptions.

This study compares wavelet transform and short-time Fourier transform for analyzing electroretinographic oscillatory potentials, finding wavelet transform provides better time-frequency resolution and disease-specific insights.

In a 63-tool benchmark, dataset identity explained ≈84% of Cohen's κ variance on synthetic data versus ≈4% for tool identity, and k-nearest-neighbor (kNN) purity predicted accuracy near-linearly (R² = 0.85–0.99).

AREION learns atlas-agnostic representations of brain connectomes, outperforming baselines with fewer parameters and strong transferability across atlases.

This paper introduces LARK, a framework that learns chemical representations from reactions, showing improved performance in reaction prediction and molecular property prediction.

This paper develops an unsupervised pipeline to predict viral fitness using sequence data and finds that protein alignment, not model size, is key for accurate prediction.

This paper introduces STRATA, a framework that models interactions between lipid nanoparticle components and their ratios to improve prediction of transfection efficiency.

TRIDENT-2 predicts chemical toxicity across eukaryotic species using multimodal transformers, achieving accuracy across diverse chemicals and taxa.

EpiZoo is a cross-species single-cell epigenomic foundation model that uses DNA sequence-aware architecture to improve cell type annotation and mutation prioritization across species.

SpatialTRACE extends anatomical annotations from a few structures to tissue-wide maps using graph and image-based models in spatial transcriptomics and microscopy.

RNASeek is a generative foundation model for RNA sequence design and functional prediction, using reinforcement learning to optimize RNA properties like stability and cleavage activity.

This paper introduces TMCA, a multimodal framework that integrates SMILES, molecular graphs, and 3D conformations for molecular property prediction, achieving the best performance on benchmark datasets.

A sparse mixture-of-experts DNA model conditioned on hyperbolic taxonomy embeddings learns better microbial representations than its taxonomy-free twin and comes close to Evo2-7B in probing and fine-tuning at lower compute.

OpenKinetics Predictor unifies thirteen enzyme kinetic predictors behind one interface and API; on E. coli iML1515, eight kcat predictors correlate pairwise at r 0.08 to 0.67.

The paper identifies a symbiont-derived biosurfactant that links host neuroimmune control of intestinal peristalsis with bacterial susceptibility to expulsion.