Questions about applying CellNavi to disease-associated cell-state transitions

[Jiahuan90]

@OXPHOS @wcanisay-y
Thank you for developing CellNavi. I have successfully installed the software and reproduced the tutorial results using the provided dataset and checkpoint.

I am interested in applying CellNavi to single-cell RNA-seq data containing multiple biological states, where the cells do not have known perturbation labels. I would be very grateful for your advice on several technical points:

1. In the current public implementation, both train_data and test_data appear to be Perturb-seq datasets with known perturbation labels. For an unlabeled biological dataset, is there a recommended inference workflow that directly takes state 1 and state 2 cells as input and outputs candidate driver genes?

2. Is the output driver-gene space strictly limited to the perturbation genes included in the fine-tuning dataset? In other words, can the model prioritize a gene that was not present among the perturbation labels used for fine-tuning?

3. The paper demonstrates generalization across cell types. Would it be reasonable to fine-tune CellNavi on one Perturb-seq dataset and then apply the model to biologically distinct cell types as an exploratory analysis?

4. Should the adjacency and distance gene graphs be reconstructed for each new cell type or dataset, or can the provided graph be reused when applying the model to other human cell types?

5. Can multiple Perturb-seq datasets be combined for fine-tuning? If so, would you recommend training CRISPRa, CRISPRi, and CRISPR knockout datasets separately?Any guidance or example code for inference on unlabeled state-transition data would be extremely helpful.

Thank you very much for your time.

[wcanisay-y]

Hi Jiahuan,

Thanks for your interest in our paper. Here are my comments.

1. CellNavi does support unlabeled data inference. In the paper, multiple methods are used to evaluate unannotated data, including transition scores computation in the T cell differentiation setting, GO enrichment of top-ranked predicted genes and clustering by likelihood score vectors in drug MoA setting, and known ER stress gene ranking in neurodegenerative disease setting.
2. CellNavi can only assign probability mass to genes that appear as labels during fine-tuning, since it's a supervised learning model.
3. This is directly supported by the paper, for example the generalization from HEK293FT cells to K562 cells. However, results may also depend on how biologically distinct the training and inference cell types are from each other.
4. No, the NicheNet-derived graph is a fixed prior reused across all cell types.
5. Combining multiple Perturb-seq datasets is an open question. I think it depends on if these datasets have batch effects or other technology bias. For the CRISPRa/CRISPRi/knockout question, my instinct is pooling them could confuse the model about the definition of driver genes since different technologies induce very different transcriptional signatures for the same gene. But you can definitely give a try and think of ways to improve.

[BoydenYubin]

In your tutorial, I couldn't find any content describing how cells from state 1 and state 2 are used as inputs to output candidate driver genes. I also noticed that the paper mentions a corresponding loss function for this part, but the details are missing from the tutorial. Could you please explain the application/workflow steps for this part in more detail? Thank you!