SIMULTANEOUS LEARNING FROM BULK AND SINGLE-CELL EXPRESSION DATA WITH PERCEIVER-BASED MODELS

Rafal Powalski, Albert Roethel, Joanna Krawczyk, Przemysław Pietrzak, Łukasz Smoliński, Maciej Sypetkowski, Paweł Biernat, Dariusz Plewczyński, Tomasz Jetka

Abstract:

In recent years, genomic experiments have generated vast amounts of expression
data, including bulk RNA-seq and single-cell RNA-seq datasets, spanning both
observational and perturbational studies. However, existing models have not fully
leveraged this diverse data landscape, focusing on modeling either bulk or single-
cell expression data. We present Funomics T0, the first foundation model that can
simultaneously learn from bulk RNA-seq and single-cell RNA-seq datasets from
observational and perturbational studies. The proposed Perceiver-based model
produces latent representations of the expression data that can be further used for
various downstream tasks. We evaluate our model on perturbation prediction and
tissue annotation tasks, using a comprehensive benchmark suite and demonstrating
strong performance across metrics, with Funomics T0 outperforming the State
model on multiple perturbation metrics.

Track: Main track

Keywords: Perturbation Modeling, Genomic Foundation Models, Multi-task Learning, Multi-modal Integration

TLDR: Funomics T0 is the Perceiver based foundation model that learns from both bulk and single-cell RNA-seq data from observational and perturbational studies simultaneously.

Submission Number: 30