Solaguren-Beascoa, A. | Search bioRxiv

bioRxiv2025

Multi-modal single-cell foundation models via dynamic token adaptation

AO_SCPLOWBSTRACTC_SCPLOWRecent advances in applying deep learning in genomics include DNA-language and single-cell foundation models. However, these models take only one data type as input. We introduce dynamic token adaptation and demonstrate how it combines these models to predict gene regulation at the single-cell level in different genetic contexts. Although the method is generalisable, we focus on an illustrative example by training an adapter from DNA-sequence embeddings to a single-cell foundation models token embedding space. As a qualitative evaluation, we assess the impact of DNA sequence changes on the models learned gene regulatory networks by mutating the transcriptional start site of the transcription factor GATA4 in silico, observing predicted expression changes in its target genes in fetal cardiomyocytes.

bioinformatics↗

bioRxiv2025

Detecting cell-level transcriptomic changes of Perturb-seq using Contrastive Fine-tuning of Single-Cell Foundation Models

Genome-scale perturbation cell atlases are an exciting new resource to understand the transcriptomic and phenotypic impact of single-gene activation or knockdown. However, in terms of differentially expressed genes identified, the signal detected in these data atlases is low, leading to the exclusion of most data from downstream analyses. Recent advances in single-cell foundation models have shown promise in capturing complex biological insights. However, their application to perturbation analysis, especially in predicting perturbed single-cell transcriptomes, remains limited. In this paper, we focus on learning representations of single-cell transcriptomes that capture subtle, yet important, transcriptome-wide changes, and we propose a novel fine-tuning strategy using contrastive learning to leverage single-cell foundation models for this task. We pre-train a single-cell foundation model and fine-tune on a genome-scale perturbation dataset using a contrastive loss, which minimises the distance between cell embeddings from unperturbed cells while maximising the distance between perturbed and unperturbed cells. We validate and test the model on unseen perturbations, demonstrating its ability to identify global biologically meaningful transcriptional changes not captured by traditional differential expression methods. Our approach provides a novel framework for analysing single-cell perturbation data and offers a more effective means of identifying perturbations that drive systemic gene expression changes.

bioinformatics↗