contrastive learning

Unlocking biological insight from single-cell data with an interpretable dual-stream foundation model

This paper introduces scDMC, a dual-stream foundation model for single-cell data that improves representation learning and achieves state-of-the-art results in cell annotation, clustering, and integration.

Contrastive learning of adverse events to provide effective and interpretable vector representations for machine-assisted pharmacovigilance

This paper develops interpretable adverse event vector representations using contrastive learning for pharmacovigilance, outperforming existing methods in drug-event causality prediction.

Flex-sweep 2.0: more flexible and faster selective sweeps detection

Flex-sweep 2.0 improves selective sweep detection by enhancing speed, memory efficiency, and flexibility with new features like custom CNN architectures and domain-adaptive training.

Learning with Recurrence Geometric AI in Spatial Transcriptomics

This paper introduces RGAI, a geometric deep-learning framework for cross-domain analysis of spatial transcriptomics, enabling alignment of latent tissue states across diverse datasets.

Charting the small-molecule universe from mass spectra with neuro-symbolic AI

AIMe uses neuro-symbolic AI to predict and annotate small-molecule spectra, expanding the searchable chemical space by three orders of magnitude.

RVQ-Alpha: Bridging Single-Cell Transcriptomics and Large Language Models via Hierarchical Discrete Tokenization and Fact-Aware Reinforcement Learning

RVQ-Alpha adapts large language models for single-cell analysis by integrating hierarchical tokenization and reinforcement learning, achieving performance comparable to task-specific models on a multi-task benchmark.

Contrastive Regulatory Embeddings Attention Model for Differential Expression Prediction with Personalized Genomes: Insights and Challenges

This paper introduces CREAM, a deep learning model for differential expression prediction using personalized genomes, showing improved performance on training genes but challenges in generalizing to unseen genes.

Contrastive learning unites sequence and structure in a global representation of protein space

This paper introduces CLSS, a contrastive learning model that co-embeds protein sequences and structures, revealing evolutionary relationships and structural transitions.

PG-MLD: Physics-Guided Molecular Representation Learning via Dynamic 3D Trajectory Distillation

PG-MLD transfers dynamic 3D physical knowledge to SMILES-based molecular models, improving property prediction without requiring 3D conformer generation.

ConfDock: Atom-specific Uncertainty Quantification for Molecular Docking via Conformal Prediction

ConfDock provides atom-specific uncertainty quantification for molecular docking using conformal prediction, achieving narrower intervals while maintaining coverage guarantees across diverse protein-ligand systems.

When Multimodal Fusion Fails: Contrastive Alignment as a Necessary Stabilizer for TCR--Peptide Binding Prediction

The paper shows that naive fusion of sequence and graph modalities in TCR-peptide binding prediction can degrade performance and introduces TRACE, a contrastive alignment method that stabilizes multimodal learning.

PlumageParts: A fine-grained avian plumage segmentation dataset and benchmark for ecological image analysis

PlumageParts provides a dataset of 4,705 bird images annotated for nine plumage regions and benchmarks segmentation models, achieving 84.01% mean Intersection over Union with a frozen DINOv3 encoder.

CENO: A Genome-Scale World Model for Evolutionary Sequence Interpretation and Programmable Regulatory Design

CENO is a long-context generative model for genomic sequence understanding and design, showing improved variant prediction and regulatory sequence generation across species.

Long-Horizon associative learning as a unifying framework for statistical learning across scales

This paper introduces a unifying framework for statistical learning across temporal scales, showing a single mechanism can capture both local and higher-order patterns in data.

DCGAN-Based Synthetic MRI Augmentation for Data-centric Brain Tumor Segmentation

This paper uses DCGAN to generate synthetic MRI images for brain tumor segmentation, improving Dice coefficient and IoU metrics compared to a baseline model.

Discovery of a phenazine thiol conjugase from sparse data using genome-informed machine learning

The paper uses genome-informed machine learning to discover a new enzyme, PTC, that catalyzes phenazine thioconjugation from sparse data.

Machine-learning MRI stratification of genetic frontotemporal dementia for clinical trial enrichment

This paper uses machine learning to stratify genetic frontotemporal dementia patients into subtypes and provides progression scores that improve trial enrichment efficiency.

HiC2Self: self-supervised denoising for bulk and single-cell Hi-C contact maps

HiC2Self denoises Hi-C contact maps using self-supervised learning, reconstructing 3D chromatin structures from low-coverage data across bulk and single-cell samples.

SpatialJEPA: JEPA-inspired graph-context distillation for spatially aware multiomics integration

SpatialJEPA transfers spatial context from spatial multiomics data to dissociated non-spatial data through graph-context distillation, improving alignment and recovery of spatially organized programs in mouse brain multiomics.

Binary node clustering via contrastive learning for haplotype phasing in de novo genome assembly

This paper introduces grapHiC, a contrastive learning framework for reference-free haplotype phasing in de novo genome assembly, achieving results comparable to state-of-the-art methods.

PepCL: A replay-based continual learning framework for updating peptide-MHC models

PepCL updates peptide-MHC predictors with new assay data while preserving prior MS knowledge, improving prediction across alleles in clinical contexts.

CAFE: A Co-folding Approach for Fragment Exploration of Allosteric and Cryptic Binding Sites

CAFE improves co-folding models for discovering allosteric and cryptic binding sites by using orthosteric blockers to guide fragment exploration.

Multicenter self-supervised computational pathology identifies prognostic histomorphological phenotypes in colorectal cancer

This paper develops a self-supervised computational pathology framework to identify prognostic histomorphological phenotypes in colorectal cancer using H&E images and tile embeddings.

Global StationaryOT: Trajectory inference for aging time courses of single-cell snapshots

This paper introduces a method to infer cell trajectories from aging time courses, improving accuracy in noisy data by leveraging temporal information.

multiScaleAC: Cell-Cell interaction with Moran's I as a function of kernel bandwidth

The paper introduces multiScaleAC, a method to analyze cell-cell interactions across spatial scales using a modified Moran's I with varying kernel bandwidths, showing improved statistical power and reduced bias.

SST-MAE: Learning Spectral-Spatio-Temporal Representations from Plant Hyperspectral Time Series to Discover Complex Genotype-Phenotype Relations

Reliability-weighted target prioritization in CD4+ T-cell Perturb-seq: a generalizability-theory decomposition

The paper introduces a method to prioritize targets in Perturb-seq by assessing the dependability of perturbation effects using generalizability theory, improving target identification reliability.

Evidence of predictive information compression in latent space in humans during speech listening

This study compares speech processing models to EEG data, finding that predictive information encoding in latent space better explains neural activity than compression or prediction alone.

Hidden sampling biases inflate performance in gene regulatory network inference

The paper shows that sampling biases in negative regulatory interactions inflate performance in gene regulatory network inference and proposes a degree-aware sampling approach to address this issue.

MolMAE: A Surface-Centric Multimodal Masked Autoencoder for Molecular Representation Learning

MolMAE is a multimodal masked autoencoder that learns molecular representations by integrating surface point clouds, 3D graphs, and SMILES-derived tokens, achieving strong performance on molecular property prediction tasks.