Question about inference peaks embedding with GET model

[flysheeeep]

Hi Xi,

Thanks for your excellent work. I’m trying to use the GET model to extract peak embeddings from ENCODE data, and I noticed that there are multiple checkpoint-best.pth files available in the AWS S3 repository.

$ aws s3 ls s3://2023-get-xf2217/get_demo/checkpoints/ --no-sign-request      
 PRE erythroblast_benchmark/                                                                                                                   
 PRE fetal_pretrain_fetal_finetune/                                                                                                            
 PRE k562_crispri_benchmark/                                                                                                                   
 PRE regulatory_inference_checkpoint_fetal_adult/  

Could you please advise which checkpoint would be most suitable for this use case, or how to choose an appropriate checkpoint for ENCODE peak data?

Thank you very much for your help.

Best regards
Feiyang

[fuxialexander]

We didn’t trained on ENCODE data, except for the K562 one. If you specifically want to focus on K562 you can use that one maybe, but otherwise I’d suggest to retrain it on e.g. ENCODE CCRE

[flysheeeep]

Got it—thanks for your reply. One point I’m still a bit confused about: you mentioned that the checkpoint
regulatory_inference_checkpoint_fetal_adult/finetune_fetal_adult_leaveout_astrocyte/checkpoint-best.pth
was trained with binary ATAC signals. In that case, could you clarify which checkpoint(s) were trained using quantitative ATAC signals?

Thanks again for the clarification.
Best,
Feiyang