search

Rappsilber-Laboratory / AlphaLink2

AlphaLink2: Integrating crosslinking MS data into Uni-Fold-Multimer
Creative Commons Attribution 4.0 International
46 stars 14 forks source link

Question about model confidence parameters when generating more than 1 model #16

Closed andreadiianni closed 9 months ago

andreadiianni commented 10 months ago

To whom it may concern, firstly many thanks for releasing this interesting tool. I would like to know how it is possible to recover interface predicted template modeling score (ipTM) when more than 1 model (e.g. 3 best models) are generated. The second question is about model ranking. What is the scoring rationale behind? In other words, when generating more than 1 model, the tool generates three models (97923, 44732 and 99982). I think the three models have been generated with different weights, and information about predicted template modeling score (pTM) of the three models can be retrieved from the notebook. I noticed that not always the model that has the highest pTM is the one that is ranked as top 1 pose (but it is fine because pTM contributes 20% to the model confidence). Then I guess the model scoring is driven by ipTM. It is correct or not? Should it be added some additional script from the colab notebook in order to recover these values when more than 1 model is generated? Many thanks again.

lhatsk commented 10 months ago

97923, 44732 and 99982 correspond to the different seeds that have been used for the prediction. The network weights will be the same.

We choose the best model in the end based on model confidence which is a compound score: 0.2 pTM + 0.8ipTM.

_ptm.json actually contains the model confidence. I fixed this and added ipTMs to the output. See AlphaLink2_iptm.json

andreadiianni commented 10 months ago

Many thanks for answering and for the implementation. Now it is possible to retrieve all different parameters for the generated models.