to_do_list.md

                                        First trial

• Macro steps:

  • ML part:
    • There is a huge work of organizing the data first. It is necessary to finally get 680 rows with each row being a complex. Features should be meaningful for the algorithm.
    • Once the df is ready with 680 features, data augmentation should be performed to let the algorithm generelize better
  • DL part:
    • Try to treat the protein and ligand as human text with NLP (look at tokenize, encoder, word embedding, text similarities, TF-IDF, word embedding)

• Next steps :

  • Download PDBBind data : done
  • Check what is molecular docking and write a slide on it : done
  • Think about how to organize the data with the different data I have :
    • try to concatenate the protein df with the ligands df horizontally
    • think about choosing the data coming from MD calculations or initial structures
  • Data structuration:
    • Build two dataframes with 680 rows each:
      • one based on _WAT.pdb files organized with protein features next to ligand features
      • on based on _WAT.pdb, protein.pdb and .mol2 files to get individual protein features next to individual ligand features next to complex features
  • Deal with missing values within df_wat_total and df_ligands_bonds
  • Choose the metric to use (check litterature)
  • Perform feature engineering on df_lig_bonds to calculate the number of bond_types in each ligand : v
  • Try to perform word embedding to give a direction from an atom to another within each ligand (read the paper)
  • Perform some feature engineering to calculate distances between atoms and complexes on df_wat_total (check biopandas and start with df_lig_bonds)
  • Transform df_wat_total to get protein and ligands features (df with 680 rows and x columns)
  • Do some feature engineering on df_lig_bonds and join the aggregated form of this df with the aggregated form of df_wat_total

• Tasks done:

  • Downloaded PDBBind data

  • Created 5 different dataframes :

    • one with 3D structures of proteins
    • one with 3D structures of ligands
    • one with bond structures of ligands
    • one with 3D structures of complexes surrounded by water molecules (which were deleted subsequently)
    • one with the dissociation rate constants

  • Creation of unique keys to make joins simpler:

    • protein_id ranging from 0 to 679
    • ligand_id ranging from 0 to 679
    • complex_id ranging from 0 to 679

  • Deleting unuseful columns / Dealing with missing values :

    • Deletion:
      • blank_1,2,3,4 ==> total empty column
      • alt_loc, insertion, segment_id ==> total empty column
      • charge ==> too much missing value, 98% even if I take ligands charges from df_lig_atoms

  • Choice of the metric ?

  • Feature engineering :

    • Ligands:
      • Number of bonds (with respect to bond types) within each ligand organised in dataframes within 680 rows to make the join easy with df_wat_total
      • Average number of bonds (with respect to bond types) within each ligand organised in dataframes within 680 rows to make the join easy with df_wat_total
      • Difference between number of bonds and the mean for each ligand
      • Average distances between atoms within a ligand
      • Total charge by ligand + average charge for each atom within a ligand
    • Proteins:
      • Average distances between atoms within a protein
      • Not taking into account b_factor and occupancy features bc always the same value

  • Related to the CASF-2016 paper:

    • Found the same values than the table S4 in the SI for power scoring using miniconda command line

                                    Second trial: for the first trial, I used wrong data 
      

---

General steps performed

---

• Train set: Refined 2020 dataset
• Test set: CASF-2016 coreset

1- Remove the comon complexes between the train and test from the train set to avoid data leakage : done

2- Need to perform clustering based on sequence alignment or structures to decrease the bias within the data
- Sequence clustering has been performed using CD-HIT with a 90% cutoff and I got a total of 1591 clusters with different size

3- Generate 3 different datasets by choosing from each cluster 3, 5 and 10 complexes with different binding affinity constants distributions
- dataset 1: choose clusters with minimum, median and maximum Kd
- dataset 2: choose clusters with minimum, 25th centile, median, 75th centile and maximum kd
- dataset 3: choose the 10 clusters based going from minimum to maximum Kds breaking down into 10 centiles
The idea being to keep 3 datasets with different binding affinity distributions, to test each of those against the ML algorithms and compare the results to a baseline we'd define to find out which dataset is less biased and would therefore give more reliable results.

4- Perform feature engineering on the 3 different datasets and create 3 final datasets - Be careful to perform the same feature engineering on the protein dataframes for the train and test

5- Perform feature selection if doable

6- Split the datasets into train/validation datasets and find the best CV (do not perform as not enough rows)

7- Try different algorithms with a baseline score to compare different algorithms (check which algo are used in the litterature and which one to use when small number of rows are used)

8- Fine tune the best algorithm

---

Feature engineering performed

---

• df_lig_bonds: (no preprocessing)
  • nb of simple, double, ar, am bonds with their means and differences
• df_ligand_atoms:
  • lig_distances: calculation of euclidian distances between atoms from the origin of the orthogonal axis (preprocessing: shift of the values to get minimum to 1 + log)
  • charges: (no preprocessing)
    • sum
    • median
    • mean
    • difference between sum and mean
    • difference between sum and median
  • atom_type (no preprocessing):
    • nb and average nb of atom types for each ligand
• df_proteins:
  • feature eng on residue_name ==> number of residues by proteins and number of residues by proteins - average number of residues(and median)
  • calculate the distances between atomic proteins with x,y and z
  • feature eng on occupancy (check EDA, no preprocessing) ==> check the meaning of the feature on pdb files
    • https://proteinstructures.com/structure/protein-databank/#:~:text=As%20mentioned%20above%2C%20they%20describe,This%20is%20called%20occupancy.
  • feature eng on b_factor (check EDA,no preprocessing) ==> check the meaning of the feature on pdb files
    • The term B-factor, sometimes called the Debye-Waller factor, temperature factor, or atomic displacement parameter, is used in protein crystallography to describe the attenuation of X-ray or neutron scattering caused by thermal motion
    • https://sci-hub.hkvisa.net/10.1021/acs.chemrev.8b00290
• Add noisy feature to perform feature selection

---

Results

---

These results come from the pythons scipt given with the CASF-2016 paper:

Date	Features	Models	Hyperparameters	Performance
30/10/2022	No feature selection	linear regresion	Default	Dataset 1: N: 245 Peason's: 0.198 RMSE: 2.15
				Dataset 2: N: 266 Peason's: 0.318 RMSE: 2.04
				Dataset 3: N: 247 Peason's: 0.219 RMSE: 2.16
				Average: N: 97 Peason's: 0.117 RMSE: 2.04
01/11/2022	Feat sel with RFE	linear regresion	Default	Dataset 1: N: 279 Peason's: 0.134 RMSE: 2.15
				Dataset 2: N: 265 Peason's: 0.307 RMSE: 2.04
				Dataset 3: N: 267 Peason's: 0.106 RMSE: 2.14
				Average: N: 267 Peason's: 0.305 RMSE: 2.05
				==> Dataset 2 gives the best results, the remaining will be performed using the latter
07	/11/2022	Feat sel with RFE with noise feature	linear regresion	Default