Theofanis Karaletsos

A Cross-Species Generative Cell Atlas Across 1.5 Billion Years of Evolution: The TranscriptFormer Single-cell Model

The paper introduces TranscriptFormer, a generative model trained on 112 million cells across 12 species, demonstrating cross-species cell type classification and disease state prediction.

Species-specific small models for cell type classification approach the performance of large single cell foundation models

The paper introduces CytoType and ESM-CE, species-specific small models for cell type classification that achieve performance comparable to large foundation models with fewer parameters.

VariantFormer: A hierarchical transformer integrating DNA sequences with genetic variations and regulatory landscapes for personalized gene expression prediction

VariantFormer predicts gene expression from diploid genomes, outperforming existing models and improving eQTL effect size estimation across ancestries and tissues.

rbio1-training scientific reasoning LLMs with biological world models as soft verifiers

This paper introduces rbio1, a biology reasoning model trained using soft verification via biological world models, achieving state-of-the-art performance on perturbation prediction.

GREmLN: A Cellular Graph Structure Aware Transcriptomics Foundation Model

GREmLN is a foundation model that uses graph structure to improve gene embeddings in single-cell transcriptomics, outperforming baselines in cell type annotation and perturbation prediction.

Pitfalls in performing genome-wide association studies on ratio traits

This paper shows that genome-wide association studies on ratio traits can be misleading, as associations may be driven by the denominator rather than the numerator, and proposes a method to correct for collider bias.

Bayesian Aggregation of Multiple Annotations Enhances Rare Variant Association Testing

BayesRVAT improves rare variant association testing by integrating multiple annotations, increasing power and detecting novel gene-disease associations in real data.

SubCell: Proteome-aware vision foundation models for microscopy capture single-cell biology

SubCell is a deep learning model for fluorescence microscopy that captures cellular morphology and protein organization, outperforming existing methods and enabling a multimodal approach to gene function.

scGenePT: Is language all you need for modeling single-cell perturbations?

The paper shows that combining language-based gene representations with experimental data improves perturbation prediction in single-cell data.

Deep Learning Analysis on Images of iPSC-derived Motor Neurons Carrying fALS-genetics Reveals Disease-Relevant Phenotypes

The paper uses deep learning on images of iPSC-derived motor neurons with fALS mutations to identify disease-relevant phenotypes and molecular differences.

EmbedGEM: A framework to evaluate the utility of embeddings for genetic discovery

EmbedGEM evaluates embeddings for genetic discovery by assessing heritability and disease relevance using polygenic risk scores and genome-wide association statistics.

An allelic series rare variant association test for candidate gene discovery

The paper introduces COAST, a rare variant association test for allelic series, which improves power over SKAT-O in detecting gene-trait associations in lipid and cell count traits using UK Biobank data.