Gunnar Rätsch

Probabilistic uncertainty in virtual spatial transcriptomics from histology is largely redundant with the predicted mean

This study benchmarks uncertainty estimation in virtual spatial transcriptomics, finding that predicted means largely explain prediction error, and uncertainty provides limited additional information.

Slide-level batch structure limits histology-guided supervision of transcriptomic foundation models

This paper investigates whether histology can improve transcriptomic foundation models by providing cross-modal supervision, but finds limited benefits due to slide-specific structure in gene-expression embeddings.

Transcriptomics-Conditioned Virtual Tissue Synthesis via Diffusion Transformers

The paper introduces STMDiT, a diffusion transformer that synthesizes H&E histopathology images from transcriptomic data, improving image quality and transcriptomic fidelity over existing methods.

Striping artifact removal in VisiumHD data through nuclear counts modeling

The paper introduces a statistical destriping method for VisiumHD spatial transcriptomics data using nuclear counts modeling, improving stripe-factor estimation and preserving biological signal.

Towards Cross-Sample Alignment for Multi-Modal Representation Learning in Spatial Transcriptomics

This paper introduces a framework that aligns spatial transcriptomics data across multiple samples by integrating transcriptomics correction with deep learning, improving integration performance over existing methods.

Perturbation-aware representation learning for in vivo genetic screens

This paper develops a method to learn perturbation-aware representations from CRISPR screens using single-cell transcriptomics, improving the identification of perturbation effects on gene expression.

Indexing All Life's Known Biological Sequences

The paper develops a scalable indexing framework for biological sequences, enabling efficient and cost-effective full-text search across vast genomic datasets.

DeepSpot2Cell: Predicting Virtual Single-Cell Spatial Transcriptomics from H&E images using Spot-Level Supervision

DeepSpot2Cell predicts virtual single-cell gene expression from H&E images using spot-level supervision, outperforming previous methods on a new benchmark.

Splicing neoepitope prediction is sensitive to methodological differences

The paper evaluates the impact of methodological differences on splicing neoepitope prediction, finding high variability and limited reproducibility across pipelines.

Molecularly informed analysis of histopathology images using natural language

SpotWhisperer links histopathology images to spatial gene expression and text annotations using natural language AI, outperforming existing models on a new benchmark dataset.

DeepSpot: Leveraging Spatial Context for Enhanced Spatial Transcriptomics Prediction from H&E Images

DeepSpot predicts spatial transcriptomics from H&E images, improving gene correlations across cancer datasets and generating virtual spatial transcriptomics samples.

SimReadUntil for Benchmarking Selective Sequencing Algorithms on ONT Devices

SimReadUntil benchmarks selective sequencing algorithms on ONT devices by simulating ReadUntil API behavior with real or simulated reads.

Learning Single-Cell Perturbation Responses using Neural Optimal Transport

This paper introduces a method to predict single-cell perturbation responses using neural optimal transport, showing improved performance over existing methods on drug-treated cells.

ResMiCo: increasing the quality of metagenome-assembled genomes with deep learning

ResMiCo improves metagenome assembly quality by accurately identifying misassembled contigs using deep learning, achieving better performance than existing methods.

Aligning Distant Sequences to Graphs using Long Seed Sketches

The paper introduces a long seed sketch method for sequence-to-graph alignment, improving time-accuracy trade-offs in high mutation rate scenarios.

PipeIT2: A tumor-only somatic variant calling workflow for Molecular Diagnostic Ion Torrent sequencing data

PipeIT2 is a somatic variant calling workflow for Ion Torrent sequencing data that identifies mutations without germline control, achieving >95% recall for variants with >10% allele fraction.

Label-guided seed-chain-extend alignment on annotated De Bruijn graphs

This paper introduces a new alignment method for annotated De Bruijn graphs that improves taxonomic classification accuracy and coverage for long reads.

SECEDO: SNV-based subclone detection using ultra-low coverage single-cell DNA sequencing

SECEDO detects subclones in tumors using SNVs from ultra-low coverage single-cell DNA sequencing, achieving high accuracy in synthetic and real datasets.

scROSHI - robust supervised hierarchical identification of single cells

scROSHI identifies cell types without training data by using pre-existing gene sets and hierarchical relationships, outperforming methods in low-data and high-diversity scenarios.

Lossless Indexing with Counting de Bruijn Graphs

This paper introduces Counting de Bruijn graphs for efficiently indexing and querying sequencing data, achieving smaller storage and faster construction compared to existing methods.

Topology-based Sparsification of Graph Annotations

The paper introduces RowDiff, a method to compactly store graph labels by exploiting annotation similarities, achieving significant space reductions when combined with compression techniques.

pmVAE: Learning Interpretable Single-Cell Representations with Pathway Modules

pmVAE learns interpretable single-cell representations by incorporating pathway structure, improving detection of perturbation-targeted pathways in single-cell RNA-seq data.

Reconstructing tumor evolutionary histories and clone trees in polynomial-time with SubMARine

This paper introduces SubMARine, a polynomial-time algorithm for reconstructing tumor clone trees that efficiently captures uncertainty and scales to large datasets.

Fast Alignment-Free Similarity Estimation By Tensor Sketching

The paper introduces Tensor Sketch, a method for fast and accurate edit distance estimation in biological sequences, outperforming existing methods in speed and accuracy.

SCIM: Universal Single-Cell Matching with Unpaired Feature Sets

SCIM matches single-cell measurements across technologies by learning a technology-invariant latent space and applying bipartite matching, achieving high accuracy in real-world datasets.

Whole genome and RNA sequencing of 1,220 cancers reveals hundreds of genes deregulated by rearrangement of cis-regulatory elements

The paper identifies hundreds of genes deregulated by SVs in cancer, showing increased expression linked to SV breakpoints and enhancer elements.

AStarix: Fast and Optimal Sequence-to-Graph Alignment

AStarix provides a fast and optimal algorithm for aligning sequences to genome graphs, outperforming existing methods by 1-2 orders of magnitude.

Sparse Binary Relation Representations for Genome Graph Annotation

This paper introduces a new compression method for de Bruijn graph labels, showing improved performance over existing methods on real-world datasets.

A Pan-Cancer Transcriptome Analysis Reveals Pervasive Regulation through Tumor-Associated Alternative Promoters

This study identifies alternative promoters as key regulators in the cancer transcriptome using RNA-Seq data from 1,188 cancer samples.

A Bayesian Nonparametric Approach to Discover Clinico-Genetic Associations across Cancer Types

This paper introduces a Bayesian nonparametric method to find genetic-clinical associations across cancer types, showing it recovers known associations and outperforms classical methods.