RosettaCommons
rf3 produced very different ipTM with or without MSA specified for the test example 5vht
For the test example 5vht, when running rf3 prediction from the json file with MSA specified or the cif file, i got ipTM of 0.90 ; however, when I ran rf3 from the json file without MSA, I got ipTM only 0.44 (python models/rf3/src/rf3/inference.py inputs='models/rf3/tests/data/5vht_from_json_noMSA.json' ckpt_path='checkpoints/rf3_foundry_01_24_latest.ckpt' out_dir='models/rf3/tests/data/5vht_pred_noMSA/').
Where could it be wrong in my setup? any suggestions are appreciated.
also interestingly, i only need to run a few designs (eg. less than 10), i can see some of the binder designs showed ipTM of greater than 0.80 when predicted by alphafold3. is this high in silico success rate is normal for rfd3/mpnn? with older version rfdiffusion pipeline (eg. the dl_binder_design), i typically need to run over 1000s designs.
Will let @Ubiquinone-dot comment on the higher in-silico success; in general, the difference in-silico success is quite target-dependent, but I'm not necessarily surprised that you are seeing much better results. For the test example 5vht, that is a native protein from the PDB. As such, I would expect it would only fold correctly when provided with the MSA; the confidence is thus working correctly (and the fold without MSA is probably wrong). Or am I misunderstanding?
@nscorley Thank you very much. I ask the question about the example 5vht, because I saw in the README.md "For this example, the pTM in the metrics.csv should be >0.8 (even without an MSA); if not, there may be something wrong with your setup."
High success rates should ofc be accompanied by diversity too but yes they should be improved compared to RFD1 - if you want more diversity (but lower success rate) step_scale<1.5 can be useful to tune to get more passing designs
