Adding Binding residues (Bind) splits

[joaquimgomez]

Once applied, this PR will add Binding residues (Bind) splits.

Proposed splits:

• Working at protein-level (train on sequences with X ligan type(s), test on else (~18% of the proteins have >1 type of ligand)):

  • one_vs_many: train on proteins with only 1 type of ligand, test on proteins with 2 and 3 types of ligands
  • two_vs_many: train on proteins with 1 or 2 types of ligand, test on proteins with 3 types of ligand
  • from_publication: train on proteins with 1, 2 or 3 types of ligand from original training set, test on original test sets TestSet300 and TestSetNew46 (mixed).

• Working at residue-level (train on sequences with residues assigned to only 1 type of ligand, test on sequences with residues assigned to multiple classes (~4% of the residues have more than one type of ligand)):

  • one_vs_sm: train on proteins with residues having only one type of ligand, test on proteins with residues having Small+Metal ligands
  • one_vs_mn: train as one_vs_sm but with balances classes, test on proteins with residues having Metal+Nuclear
  • one_vs_sn: train as one_vs_sm but with balances classes, test on proteins with residues having Small+Nuclear

The splits one_vs_many, two_vs_many and from_publication aim to analyze the impact of training/testing taking into account that proteins have more than one type of ligands, i.e., we can have proteins with only residues with metal ligands and/or proteins with residues with metal ligand and other residues with small ligand.

On the other hand, the splits one_vs_sm, one_vs_mn and one_vs_sn aim to analyze the impact of training in proteins with residues having only one type of residue and testing on proteins that hve multi-ligand residues.

To do:

• [x] Upload to GitHub final splits files
• [x] Add README
• [x] Add splits to FLIP/splits/README.md
• [ ] Add original data to the server (?)

Open issues:

• [ ] Impossible to create statistics for the splits because the target is not a single value.
• [ ] Residues indexes start with 0? Implemented considering this.

[sacdallago]

three_vs_many: train on proteins with 1 or 2 types of ligand from original training set, test on original test sets TestSet300 and TestSetNew46 mixed.

Why is it called three_vs_many?

[joaquimgomez]

three_vs_many: train on proteins with 1 or 2 types of ligand from original training set, test on original test sets TestSet300 and TestSetNew46 mixed.

Why is it called three_vs_many?

There was a typo. "with 1 or 2 types of ligand" ==> "with 1, 2 or 3 types of ligand". If you prefer another name, I am open to suggestions. Maybe all or many_vs_many?

[sacdallago]

Some observations:

1. It would be beneficial to add a separate section with a legend in the README for the types (instead of just having it in the target).
2. three_vs_many and one_vs_many encode the target differently (a stringified array vs. a string).
3. If I understand correctly, the three_vs_many split is the same as in the publication? Then just call it "from_publication" or something :) makes it easier to know: OK, this is the one that I can compare to whatever table is in the manuscript.
4. This is a residue to class problem, as such, CSV is probably not the best way to encode the data. You can use the standard we developed in bio-trainer: https://github.com/sacdallago/biotrainer/blob/main/data_standardization.md#residue---class . This translates to: you can have a single "master" sequence file (simply call it sequences.fasta), and then you can have as many label files as you have splits

[joaquimgomez]

Some observations:

1. It would be beneficial to add a separate section with a legend in the README for the types (instead of just having it in the target).

-> Legend section added to the README.

2. three_vs_many and one_vs_many encode the target differently (a stringified array vs. a string).

-> Solved.

3. If I understand correctly, the three_vs_many split is the same as in the publication? Then just call it "from_publication" or something :) makes it easier to know: OK, this is the one that I can compare to whatever table is in the manuscript.

-> Name changed to "from_publication" everywhere.

4. This is a residue to class problem, as such, CSV is probably not the best way to encode the data. You can use the standard we developed in bio-trainer: https://github.com/sacdallago/biotrainer/blob/main/data_standardization.md#residue---class . This translates to: you can have a single "master" sequence file (simply call it sequences.fasta), and then you can have as many label files as you have splits

-> Splits files changed accordingly with the standardization in bio-trainer.

[joaquimgomez]

I double-checked all the files concerning the PR. I think they are ready for a merge if there are no more comments.

[sacdallago]

Great!