search

Freshwind-Bioinformatics / TABR-BERT

TABR-BERT: an Accurate and Robust BERT-based Transfer Learning Model for TCR-pMHC Interaction Prediction
Other
8 stars 1 forks source link

too high rank value #1

Closed zunyun-Gong closed 7 months ago

zunyun-Gong commented 7 months ago

Thank you for sharing your fasinating work. When I use it, I find all my TCR-pMHC get a very high rank which in your paper represent high possibility for binding. But, in our perspective it can be possible because there all both selected randomly from tumor T cell repertoire and neoantigens. Is there anything wrong?

c7ee1d2c0a5f54e1b75cdf46cef2b95

code:singularity exec --pwd /mnt -c --bind /tools/TABR-BERT/:/mnt,/mydata/tabr/:/boot tabr-bert_v1.sif python /workspace/TABR-BERT/predict_tcr_pmhc_binding.py --input /boot/0t0_tabr.list --healthy_tcr ./data/small_healthy_tcr.csv --pseudo_sequence_dict ./data/mhcflurry.allele_sequences_homo.csv --output /boot/test.txt --GPUs 0

JiaweiZhang1997 commented 7 months ago

This is possible because TABR-BERT returns a rank value, which suggests that the predicted TCRs have a higher probability of combining with the neoantigen relative to the TCRs in "healthy_tcr", and you might consider replacing "healthy_tcr" with your own set of TCRs to get more discriminatory results.

zunyun-Gong commented 7 months ago

Thank you for your answer! It helps me a lot.