deep learning

BoneGraph: A Domain-Specialised, Self-Correcting Reasoning System for Bone Science Retrieval, Grounded Inference, and Image Mechanics

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%.

DELPHAI predicts heterogeneous perturbation responses with learned single-cell fitness

DELPHAI predicts heterogeneous perturbation responses by jointly training a fitness network and optimal transport network to identify differentially expressed genes and depleted cell lineages.

Functional spatial transcriptomics uncover LMO7 as a fusion-regulated and clinically relevant driver of metastasis in Ewing sarcoma

The paper identifies LMO7 as a fusion-regulated driver of metastasis in Ewing sarcoma using spatial transcriptomics and integrates clinical data to show its association with poor outcomes.

Discovery of microbial intergenic features with genomic language modeling and multimodal search

The paper identifies microbial intergenic features using a sparse autoencoder and multimodal search, revealing previously uncharacterized regulatory DNA and structured RNA elements.

Time-series foundation modeling enables accurate lake ecosystem forecasting

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.

Integrating Genomic Annotations and Traits Dependencies for single-nucleotide polymorphisms Prioritization with Causal Concept Bottleneck Models

The paper introduces C2BM, a deep learning model that improves SNP prioritization by integrating genomic annotations and trait relationships, showing better predictions when related trait labels are available.

DepoCat: Interactive database of experimentally verified phage depolymerases

DepoCat is a database of experimentally verified phage depolymerases that enables sequence and structure-based classification and specificity prediction.

Integrating structural and biological evidence to rerank ESMFold2 protein-protein interactions

The paper develops a framework to rerank protein-protein interactions using structural and biological evidence, improving prediction accuracy over structural-only methods.

Predictive control of human pancreatic cell fate using a digital model of in vitro differentiation

The paper develops a predictive digital model of human pancreatic cell differentiation using single-cell multiomic data to identify regulators and validate their roles in islet fate specification.

Drought and rewetting drive divergent microbial dynamics across soil compartments and reveal Trinickia as a key genus for soybean drought resilience

This study tracks microbial dynamics during drought and rewetting in soybean soils, identifying Trinickia as a key genus linked to drought resilience.

Size Control of hnRNPK-based Nucleolar Condensates by RNA-Regulated Fusion Dynamics

This study identifies how RNA-regulated fusion dynamics of hnRNPK-based nucleolar condensates prevent coalescence and maintain size control.

Oncogenic fusions induce an extensive cancer-restricted cryptic proteome in Ewing sarcoma

The paper shows EWSR1::FLI1 fusion drives the production of cancer-specific neoproteins in Ewing sarcoma, revealing a reversible dark proteome with therapeutic potential.

High-throughput physics-based enzyme engineering

This paper introduces a physics-based framework for enzyme engineering that combines machine learning and molecular mechanics to predict catalytic rates and selectivity with high accuracy.

FORGE audits residue-level information encoded in RNA tertiary structure geometry

FORGE converts RNA geometry into interpretable descriptors to identify native nucleotides and predict base-pair states, showing high accuracy and separable properties in RNA structure analysis.

Network Modeling Predicts How DYRK1A Inhibition Promotes Cardiomyocyte Cycling after Ischemic/Reperfusion Injury

The paper develops a computational network model predicting DYRK1A inhibition promotes cardiomyocyte cycling after ischemic/reperfusion injury, validated through experiments in neonatal rat cardiomyocytes.

Standing genetic variation drives polygenic adaptation to different environmental shifts

This study uses simulations to show how environmental shifts affect genetic adaptation, highlighting the role of standing genetic variation and using a CNN to predict environmental changes and mutation effects.

A Gene-Program Architecture of Mouse T cells

The paper introduces immgenT-GP, a gene-program framework that identifies 200 reproducible gene programs in mouse T cells, capturing lineage identity, activation states, and tissue location.

Structure based discovery of antipsychotic-like TAAR1 agonists

The paper identifies TAAR1 agonists with antipsychotic-like effects through computational screening and experimental validation, showing improved potency and reduced side effects in a schizophrenia model.

Post-hoc long-read sequencing links leukemic mutation status to single-cell transcriptomes

The paper develops a post-hoc method to link mutation status to single-cell transcriptomes using long-read sequencing, showing it can detect BCR::ABL1-positive cells in chronic myeloid leukemia.

Evaluation of methods for AlphaFold-based integrative modeling

This study evaluates AlphaFold-based integrative modeling methods using chemical crosslinks, finding they often rely more on learned priors than input data.

Integrative pipeline to profile and target endocrine therapy-insensitive cell populations in ER+ breast cancer

The paper develops an analytical pipeline to identify and target endocrine therapy-insensitive cells in ER+ breast cancer using single-cell RNA sequencing and functional validation in organoid models.

Ourotide: decoding the hierarchical peptide recognition for generative design

The paper introduces Ourotide, a deep learning framework for peptide recognition that outperforms generalist models by capturing a unique hierarchical binding mode through physics-guided data augmentation.

Interaction Profiles as a Universal Language for Generative Molecular Design with ShEPhERD-2

This paper introduces ShEPhERD-2, a 3D generative model that creates drug-like molecules based on interaction profiles, demonstrating its versatility across multiple design tasks.

HaloUMI: Physics-informed analysis of inhibition halo assays

HaloUMI provides an automated tool for analyzing microbial inhibition assays, improving accuracy by accounting for irregular zones and varying microbial density.

Automatic Quality Control and Error Correction in MRI linear registration via a Residual Parameter Prediction Network for T1w MRI

A 3D ResNet18 trained on synthetically misaligned T1w scans predicts residual affine parameters, correcting synthetic misalignments to 0.778 mm residual RMSE and flagging failed registrations better than RegQCNET or DARQ.

Repeat element-anchored enhancer-promoter interactions (REEPIs) shape the regulatory landscape of hormone-resistant breast cancer

The paper identifies repeat element-anchored enhancer-promoter interactions (REEPIs) in hormone-resistant breast cancer and links them to transcriptional adaptation and regulatory reprogramming.

Divergent specializations for motion-driven representations in higher lateral and dorsal visual areas

The paper investigates how artificial neural networks trained on motion-related tasks develop representations that align with brain activity and support object categorization using motion cues.

From in silico prediction to experimental validation: Identification of drugs and novel synergistic combinations that inhibit growth of inflammatory breast cancer cells

This study validates computational drug repurposing predictions for inflammatory breast cancer using experimental assays, identifying synergistic combinations that inhibit cancer cell growth and signaling pathways.

The Unreasonable Effectiveness of Cell Types in Describing Neuronal Physiological Features

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.

Genomic foundation model-derived disruption profiling links somatic mutations to cancer biology and clinical outcomes

This paper uses genomic foundation models to assess how somatic mutations disrupt cancer biology and clinical outcomes across multiple cancer types.