Unconditional samples from the LD3 model

[Ma-Yiming]

Hi,
I run python proteinfoundation/generate.py --config_name inference_ucond_notri_long and produce 100 samples for each of the lengths in [300, 400, 500, 600, 700, 800].
However, I've noticed that the scTM and RMSD metrics for data with lengths greater than 500 are not very good. For example, the mean scRMSD for data with length 800 is 12.37. Is this due to a problem with the current script's parameter settings or with the model parameter upload?

[tgeffnerNVIDIA]

Hi Ma-Yiming,

Thank you for reaching out. This observation is not necessarily incosistent with the results reported in the paper. As we show in the third plot of Figure 4, while La-Proteina still outperforms the baselines at longer lengths, its co-designability values do decrease as the sequence length increases. You should be able to reproduce the values presented in the paper using the provided code and checkpoint, but please let us know if you encounter any issues.

Alternatively, the first plot in Figure 4 shows that using MPNN-generated sequences typically works better than the self-produced sequences for longer chains. This is indicated by the designability being higher than the co-designability at these lengths.