Applicability of TranscriptFormer to plant single-cell transcriptomic data

[liuyt57]

Thank you very much for developing and sharing TranscriptFormer. I found this work highly impressive, particularly its ability to learn cross-species representations of single-cell transcriptomes across a broad evolutionary range.

I am interested in applying this model to plant single-cell or single-nucleus RNA-seq data, particularly from crop species such as maize and rice. However, I noticed that the current models were trained on animals, yeast, and a protist, without any plant species.

I would like to ask:

1. Can the current pretrained TranscriptFormer models be directly applied to plant single-cell transcriptomic data?
2. How are genes from a new species mapped to the model vocabulary? Does the model rely on orthologous genes, pretrained protein embeddings, gene symbols, or another mapping strategy?
3. Would plant-specific genes that are absent from the current vocabulary simply be ignored?
4. Is it possible to extend the vocabulary and fine-tune or continue pretraining the model using plant species?
5. Do you have any recommendations for adapting TranscriptFormer to evolutionarily distant species such as plants?
6. Are there plans to release a plant-specific or more broadly eukaryotic version of TranscriptFormer?

More specifically, I am interested in whether the model could be used for tasks such as cross-species cell-type annotation, gene regulatory relationship prediction, and transcription factor prioritization in plants.

Thank you for your help.