contrastive learning

CLEAR-HPV: Interpretable Concept Discovery for HPV-Associated Morphology in Whole-Slide Histology

CLEAR-HPV discovers interpretable morphologic concepts in HPV-associated histology slides using attention-based methods, reducing high-dimensional features to 10 concepts while preserving predictive information.

BRIDGE-GRN: Role-Aware Bi-Tower Graph Learning with Cross-View Contrast for Directed Gene Regulatory Network Inference

BRIDGE-GRN improves directed gene regulatory network inference from scRNA-seq data by using role-aware graph learning and cross-view contrastive regularization, achieving strong performance across multiple benchmarks.

GatorDuo: Global-Consistency Dual-Graph Refinement With Pseudo-Label Agreement for Spatial Transcriptomics

GatorDuo improves spatial domain identification in spatial transcriptomics by refining dual graphs with pseudo-label agreement and adaptive clustering.

Redesign selective protein binders using contrastive decoding

RedNet improves protein binder design by incorporating side-chain information and contrastive decoding to enhance binding affinity, folding stability, and selectivity.

Pretraining Objective Shapes Cross-Category Generalization in Affective Image Prediction: A Geometric Comparison of Vision Transformer Encoders

This paper compares how different pretraining objectives affect the generalization of Vision Transformers in affective image prediction, finding that contrastive learning outperforms classification pretraining in cross-category generalization.

A Conditional Random Field approach for de novo reconstruction of bacterial haplotypes from a de Bruijn graph representation

This paper introduces HaploDetox, a method for de novo bacterial haplotype reconstruction using a Conditional Random Field approach on de Bruijn graphs, achieving improved SNP phasing accuracy.

Reliable Molecular Retrieval from Mass Spectra using Conformal Prediction

This paper applies conformal prediction to molecular retrieval from mass spectra, producing reliable, spectrum-specific candidate sets with controlled reliability.

Unified Genomic and Chemical Representations Enable Bidirectional Bio-synthetic Gene Cluster and Natural Product Retrieval

This paper introduces a multimodal framework that links biosynthetic gene clusters and natural products by integrating genomic and chemical data through shared embeddings, enabling bidirectional retrieval and strain selection.

Image-Conditioned Diffusion for Privacy-Preserving Synthetic Medical Images

This paper evaluates image-to-image diffusion for generating privacy-preserving synthetic medical images, finding it outperforms text-to-image methods in fidelity and clinical signal preservation.

DOMINO: Learning Domain Co-occurrence for Multidomain Protein Design

DOMINO generates multidomain proteins by learning domain co-occurrence patterns and using them to design novel sequences with high structural confidence.

GENERator-v2: Reconciling Coarse Tokenization with Single-Nucleotide Resolution in Genomic Language Modeling

This paper introduces improved genomic language models that balance long-range context with single-nucleotide resolution, achieving better performance and efficiency than prior methods.

DPLM: Dynamics-aware Protein Language Model via contrastive learning between sequence and molecular dynamics simulation trajectory

This paper introduces DPLM, a dynamics-aware protein language model that integrates molecular dynamics simulations with sequence data to improve predictions of protein properties.

Discovery of a Systemic Immune-Inflammatory Axis Bridging the Neuroendocrine System and Breast Tumor Microenvironment via MMD-Regularized Cross-Tissue Latent Alignment

This paper identifies a systemic immune-inflammatory axis linking neuroendocrine regulation and breast tumor microenvironment using cross-tissue latent alignment of transcriptomic data.

Molecular maps of diseases from omics data and network embeddings

The paper constructs disease maps by integrating omics data with protein network embeddings, identifying more disease-relevant proteins than omics analysis alone.

Mapping AAV capsid sequences to functions through function-guided in silico evolution

The paper introduces ALICE, a platform that uses function-guided in silico evolution to design AAV capsids with tailored properties, demonstrating improved targeting and tropism in mice.

Breast cancer radiogenomics analysis via computational perturbation on AI-driven multi-omics guided image synthesis

This study uses computational perturbation on AI-generated MRI data to identify pathway-specific radiogenomic signatures in breast cancer.

GatorSC: Multi-Scale Cell and Gene Graphs with Mixture-of-Experts Fusion for Single-Cell Transcriptomics

GatorSC improves single-cell transcriptomic analysis by integrating multi-scale cell and gene graphs with a Mixture-of-Experts fusion framework, outperforming existing methods in clustering, imputation, and annotation.

TxConformal: Controlling False Discoveries in AI-Driven Therapeutic Discovery

This paper introduces TxConformal, a framework that controls false discoveries in AI-driven therapeutic discovery by adjusting for distribution shifts and providing confidence measures for AI scores.

Multi-Scale Contextual Attention for Robust Crop and Pest Image Classification

This paper introduces a Multi-Scale Contextual Attention module for crop and pest image classification, achieving high accuracy on a field-style dataset while outperforming baselines.

Exposure to naturalistic occlusion promotes generalized, human-like robustness in deep neural networks

This study shows that training DNNs with natural occluders improves their robustness to real-world occlusion, unlike artificial occluders which do not replicate human-like performance.

Cross-modality representation and multi-sample integration of spatially resolved omics data

This paper introduces PRESENT, a contrastive learning framework for integrating spatially resolved multi-omics data across samples and modalities, showing improved performance in detecting spatially functional domains and biological regulatory mechanisms.

Mosaic integration of spatial multi-omics with SpaMosaic

SpaMosaic integrates spatial multi-omics data using contrastive learning and graph neural networks, improving spatial domain identification and missing modality imputation across diverse datasets.

Probabilistic coupling of cellular and microenvironmental heterogeneity by masked self-supervised learning

Mievformer uses masked self-supervised learning to model microenvironmental and cellular heterogeneity in spatial transcriptomics data, outperforming existing methods in niche clustering and real dataset analysis.

H2O: A Foundation Model Bridging Histopathology to Spatial Multi-Omics Profiling

H2O predicts spatial transcriptomics and proteomics from H&E images using a contrastive learning framework, outperforming existing models and recovering biologically relevant cell-cell communication.

DeepBranchAI: A Novel Cascade Workflow Enabling Accessible 3D Branching Network Segmentation

DeepBranchAI reduces annotation time for 3D branching network segmentation by using a cascade workflow that transforms sparse labels into comprehensive training data.

Evaluating Expert Specialization in Mixture-of-Experts Antibody Language Models

This study evaluates MoE architectures for antibody language models, finding token-choice routing improves performance by specializing in CDRH3 residues and enabling efficient pre-training.

An Interpretable 3D Bag-Of-Visual-Words Pipeline for Volumetric Microscopy Classification

This paper introduces an interpretable 3D Bag-of-Visual-Words pipeline for volumetric microscopy classification, showing it can accurately classify and localize biologically relevant structures in complex datasets.

PEACE: Prototype-aware Effector Analysis via Contrastive Embeddings

PEACE improves effector protein prediction by using contrastive embeddings and prototype-aware training, outperforming EffectorP 3.0 on fungal and fungi+oomycete datasets.

Blood-to-tissue translation in autoimmune disease: paired single-cell evidence from systemic sclerosis

This study shows that peripheral blood can reflect tissue-resident molecular states in systemic sclerosis through paired single-cell RNA-sequencing, with strong recovery of fibroblast subtype programs.

SingleBehavior Lab: behavioral sequencing and phenotyping with lightweight task specific adaptation

SingleBehavior Lab (SBL) models behavior across experimental contexts using spatiotemporal embeddings and lightweight adapters, enabling behavior sequencing and task-specific refinement with few-shot learning.