Generative pre-trained models are the new cool kid on the block 😎 that no one can ignore anymore. Usage of this new approach have tremendous impact in a broad amount of areas. Biology is no exception, with complex molecular signatures, different cell types, states, and all kinds of different experiments to induce alteration in specific mechanisms make this a perfect storm to get more meaningful information from this source.
In a single approach, scGPT has managed to get 6 different important aspects from scRNAseq: 1) Clustering; 2) Batch Correction; 3) Cell type annotation; 4) Gene network inference; 5) Multiomic integration; and 6) Perturbation prediction;
For this first version of scGPT, it was used 33 million cells from CELLxGENE collection with 51 organs / tissues and 441 different studies, which represented a most of cellular differences across human body.
Authors have reported that their model without any fine-tuning can achieve impressive >80% precision while annotating non-rare cell types, and with some fine-tuning this precision can reach ~85%. Looking at perturbation, removal of 1 or 2 genes, in a cell model, fine-tuned scGPT was able to reach 60% of Pearson correlation, outcoming previous models by 5-20%. Integration of datasets also outperformed known techniques from scVI, Seurat and Harmony by 5-10%. GRN-inference was also tested finding not just known networks in curated datasets but also being able to predict interference in those networks.
Although too good to be true, we definitely can´t take our eyes out from start using more and more this type of approach as Satija (Seurat main coordinator) have highlighted in his last Single Cell Genomics Day.
Conclusion is that we are reaching a point that most of manual / semi-automated work running pipelines to describe cell types + extracting meaningful information scRNAseq would be mostly done by one of those generative models. In a near future, we might ask an LLM (as GPT4) with our fastqs which cell types and meaningful biological information can be found in our scRNAseq experiments.
Authors were also very generous to provide lots of examples in jupyter notebooks in their github repo scGPT. What you think about we explore this here in our blog for a next post? - Provide us some feedback in our Linkedin post!