Simulating unseen perturbations using scGPT

Hi there,

CRISPR-Cas screening combined with scRNA-seq analysis has advanced functional genomics, though unbiased interrogation of polygenic phenotypes has remained elusive. While CRISPR-Cas systems can induce combinatorial perturbations, it would theoretically take a minimum 400 million experiments (20,000 protein coding genes * 20,000 protein coding genes) to screen every possible two-gene permutation. Assaying multiple cell types and conditions can easily push experimental numbers into the billions.

Foundation models trained on existing perturbation datasets raise the exciting possibility of performing such experiments in silico, enabling the screening of billions of perturbations computationally. The newest Superbio application fine-tunes scGPT from Cui et al. (Nature Methods, 2024) on a Perturb-seq dataset, enabling the prediction of post-perturbation gene expression.

Read on for tips on using the Superbio application, and an example of results 👇

💻 Use transfer learning to generate a new perturbation model

Fig 1a, Cui et al (2024), Nature Methods.

Superbio provides demo data from the Adamson dataset for fine-tuning, a scRNA-seq dataset containing both single and combinatorial perturbations. Users can re-train scGPT - originally developed by Cui et al. and released in Nature Methods - on this dataset, transferring the model’s understanding of single-cell transcriptomics towards predicting gene expression resulting from unseen perturbations.

To perform fine-tuning, simply navigate to Superbio and click ‘Use Demo Data’, or upload your own perturbation data here.

🧬 Identify the top 20 DEGs between perturbation conditions

Pictured above: workflow outputs from the fine-tuned scGPT perturbation application on Superbio.

After fine-tuning, users can examine the top 20 differentially expressed genes between a guide RNA of choice and a non-targeting sgRNA control - pictured about for a simulated perturbation at the SRPRB locus.

Simply launch the scGPT Perturbation workflow, let it complete, and use the visualizations to examine specific loci of interest. Launch the model here.

Want more perturbation prediction capabilities from Superbio? Let us know in the poll below:

Stay curious,
Berke from Superbio
P.S. - Want to learn about more of Superbio’s features?
Find out detailed scGPT tutorials here.