Little to no difference in Predicted binding affinity for scFv/antigen tests

[avilella]

Hi all,

I tested PPLM for a known antibody/antigen pair, in this case pembrolizumab versus human PD-1.

I did a test with the pembrolizumab wild type, inputting in the heavy variable and light variable chains together with a GGGGSx4 linker or separately, and the reported Predicted binding affinity was the same.

When I test pembrolizumab wild type, I get:

(PPLM) user@gpuserver1:~/bin$ time python ~/PPLM/run_pplm-affinity.py ~/Downloads/pd1.trim.fasta ~/Downloads/5GGS_WT.fasta
Predicted binding affinity: -11.108373

real    0m25.301s
user    4m56.369s
sys     0m19.365s
(PPLM) user@gpuserver1:~/bin$ cat ~/Downloads/5GGS_WT.fasta
>5GGS_WT_H
QVQLVQSGVEVKKPGASVKVSCKASGYTFTNYYMYWVRQAPGQGLEWMGGINPSNGGTNFNEKFKNRVTLTTDSSTTTAYMELKSLQFDDTAVYYCARRDYRFDMGFDYWGQGTTVTVSS
>5GGS_WT_L
EIVLTQSPATLSLSPGERATLSCRASKGVSTSGYSYLHWYQQKPGQAPRLLIYLASYLESGVPARFSGSGSGTDFTLTISSLEPEDFAVYYCQHSRDLPLTFGGGTKVEIK

When I do a full ablation of all 6 CDRs, changing them for ALAnines, the reported binding decreases only by 7.3%, which is not what I was expecting:

(PPLM) user@gpuserver1:~/bin$ time python ~/PPLM/run_pplm-affinity.py ~/Downloads/pd1.trim.fasta ~/Downloads/5GGS_ALA.fasta
Predicted binding affinity: -10.344137

real    0m24.009s
user    4m33.160s
sys     0m17.483s
(PPLM) user@gpuserver1:~/bin$ cat ~/Downloads/5GGS_ALA.fasta | grep -C10 A
>5GGS_WT_H
QVQLVQSGVEVKKPGASVKVSCKASAAAAAAAAMYWVRQAPGQGLEWMGGIAAAAAAANFNEKFKNRVTLTTDSSTTTAYMELKSLQFDDTAVYYCAAAAAAAAAAAAAWGQGTTVTVSS
>5GGS_WT_L
EIVLTQSPATLSLSPGERATLSCRASAAAAAAAAAALHWYQQKPGQAPRLLIYAAAYLESGVPARFSGSGSGTDFTLTISSLEPEDFAVYYCAAAAAAAAAFGGGTKVEIK

See the ALAnines highlighted in red below, corresponding to the 6 CDRs of the original pembro:

Below is the trimmed version of human PD-1 as per the sequence chosen in the PDB crystal structures:

$ cat ~/Downloads/pd1.trim.fasta
>PD1 Q15116
WNPPTFSPALLVVTEGDNATFTCSFSNTSESFVLNWYRMSPSNQTDKLAAFPEDRSQPGQDCRFRVTQLPNGRDFHMSVVRARRNDSGTYLCGAISLAPKAQIKESLRAELRVTE

Is there a way to improve the results? I was expecting PPLM to be able to differentiate between an known binder and a fully ablated version of the mAb. Am I running it with the wrong parameters? The code above is all done from git clone of master as of 24 hours ago.

Thanks in advance

[junliu621]

Hi,

Thank you for the detailed test. Your command and input format appear to be correct, and there are no additional parameters required.

PPLM-Affinity was trained to predict the overall binding affinity of a protein pair, rather than mutation-induced affinity changes. Replacing all six CDRs with alanine is an extreme perturbation that is outside the distribution represented in the training dataset. Therefore, the difference between the wild-type and CDR-ablated predictions should not be interpreted as an accurate ΔΔG estimate.

The predicted change is in the expected direction, but the model may underestimate the magnitude of the binding loss in this case. A model specifically trained for mutation effects or binding ΔΔG prediction would be more appropriate for accurately evaluating the impact of CDR substitutions on binding affinity.