Title:	Deep learning modeling of single cell perturbations
DNr:	Berzelius-2025-401
Project Type:	LiU Berzelius
Principal Investigator:	Kasper Karlsson <kasper.karlsson@ki.se>
Affiliation:	Karolinska Institutet
Duration:	2026-01-01 – 2026-07-01
Classification:	10203
Homepage:	https://ki.se/personer/kasper-karlsson
Keywords:

Abstract

Differentiation therapies are showing promise in treating neuroblastoma, one of the most common and deadly pediatric cancers, by inducing tumor cells into more benign, differentiated states. High risk neuroblastomas are treated with retinoic acid (RA), but the clinical utility is limited due to toxicity and potentially transient differentiation effects, underscoring the urgent need for novel therapeutic strategies. While alternative differentiation-inducing drugs, such as HDAC inhibitors, have shown potential, systematic exploration of gene targets that could similarly induce neuroblastoma differentiation remains largely unexplored. Leveraging recent advances in artificial intelligence techniques capable of modeling complex biological systems, this project aims to systematically identify novel gene targets capable of inducing differentiation in neuroblastoma. We will perform extensive computational in silico studies, examining both single-gene and combinatorial gene perturbations, to predict differentiation-associated cellular responses at the single-cell level. Specifically, we are adapting a state-of-the-art deep learning model to accurately represent neuroblastoma cellular states and their responses to genetic perturbations while accounting for biological and technical confounders, such as variations in cell type, cell-cycle stages, and experimental batch effects. Following these computational predictions, we will experimentally validate the top candidate genes and gene combinations using CRISPR-based screening techniques combined with single-cell transcriptional analysis. The project is ongoing, and lessons learned thus far have highlighted methodological challenges, including rigorous data preprocessing requirements, integration of heterogeneous datasets, and careful selection of analytical strategies. Insights gained from reviewing recent foundational models, such as scGPT and scFoundation, have underscored the critical need for explicit approaches to handle confounders, further guiding our current efforts. Addressing these challenges remains central to refining robust models and laying a strong foundation for subsequent experimental validations.