d4-rescore

Table of Contents

Intro
Results
Method
Refs

Intro

In [1], Lyu et al dock the Enamine REAL virtual library at the D4 receptor (5WIU). They selected ~550 ligands at random from a range of high- and low-scoring buckets, and test these in vitro at 10µM. This represents a perfect test-case for re-scoring algorithms:

• The ligands are selected randomly
• There are a lot of actives and inactives, giving relatively high-precision estimates
• The actives all bind to the same binding site, and even the same protein conformation (with reasonably high confidence), and the re-scoring algorithms have access to this protein conformation
• The inactive labels are experimentally determined
• None of the actives are congeneric series, and none are based on prior knowledge (as is often the case in ChEMBL)
• The setting is a real use-case - i.e. re-ranking docked structures to identify hits
• The receptor is an 'easy' case: a small, enclosed, polar binding site

Thus these data are perfect for rescoring - if an algorithm gets it right, we can be confident that it gets it right for the right reasons. Likewise, if it fails to distinguish actives and inactives, we can be reasonably confident there's no missing information that would have saved it.

One caveat to this is that docking algorithms have high false negative rates: they leave a majority of the hits in the noise. We only care for the ability to select useful number of hits into the enriched region. With 122 hits in this dataset, though, we can be quite confident that it's not simply an unlucky set of ligands and that another set of actives would have performed better.

results

The tested re-scoring algorithms were: PLECScore, RFScore, and NNScore (BINANA features), which are available in ODDT, as well as RF-Score-VS-v1. With a nod to the Rognan lab's paper showing re-scoring algorithms are outperformed by scoring similarity to a known ligand, I also tested RDKit's 'feature map vectors', a similarity score between pharmacophoric points [2].

In short, feature map vectors out-perform any of the other methods, followed by vanilla Smina score and/or PLECScore, which are slightly better than random depending on what metric you prefer. NNScore, RFScore, and RF-Score-VS, do not appear to recognise actives at a higher rate than inactives, even looking only at early enrichment.

This largely agrees with [3] - their GRiM technique is analogous to feature map vectors, and they show that learning from a crystallized ligand outperforms re-scoring techniques. What if there's no ligand available? Well, docking alone still performs best.

ROC:

Early enrichment metrics:

After 'preparing' the ligands, i.e. enumerating tautomers, charge states, and enantiomers, there is little change:

ROC:

Early enrichment metrics:

steps to reproduce:

1. read and embed ligands

See 1-read_and_embed_ligands.ipynb. This step takes the SMILES codes given in the supplementary of Lyu et al [1] and prepares them for docking in two ways: the first is directly embedding each ligand in 3D using RDKit's EKTDG method, the second is enumerating tautomers/charge states/enantiomers with the Durrant lab's Gypsum-DL, which also embeds into 3D. The input files are in ./data/.

2. prepare the protein for docking

This workflow uses the script available at https://github.com/ljmartin/pdb_to_pdbqt . The resulting PDB was converted to .pdbqt format with obabel proteinH.pdb -xr -O proteinH.pdbqt. This file is in ./data/.

3. dock!

Docking was run with Smina:

cd ./data/
smina -r proteinH.pdbqt -l ligands3d.sdf --autobox_ligand AQD_ligand.pdb -o ligands3d_docked.sdf
smina -r proteinH.pdbqt -l ligands3d_gypsum.sdf --autobox_ligand AQD_ligand.pdb -o ligands3d_gypsum_docked.sdf

4. re-score and evaluate

See 4-rescore_docked_mols.ipynb, which uses ODDT and the RDKit to re-score all the docked poses, and calculate early-enrichment metrics like Robust Initial Enrichment, log(area under the ROC), BEDROC, and average precision.

RF-Score-VS is not available via ODDT, but it is available as a binary from here.

re-scoring command for RF-Score-VS is:

/path-to-binary/rf-score-vs_v1/rf-score-vs --receptor ./proteinH.pdb ./ligands3d_docked.sdf -O ./rfscorevs.csv
/path-to-binary/rf-score-vs_v1/rf-score-vs --receptor ./proteinH.pdb ./ligands3d_gypsum_docked.sdf -O ./rfscorevs_gypsum.csv

Refs

[1] Ultra-large library docking for discovering new chemotypes Lyu et al.

[2] Feature-map vectors: a new class of informative descriptors for computational drug discovery, Landrum et al.

[3] True Accuracy of Fast Scoring Functions to Predict High-Throughput Screening Data from Docking Poses: The Simpler the Better Tran-Nguyen et al.