NbX

This repository contains the NbX models for the re-ranking of nanobody–antigen binding poses.

Tam, C.; Kumar, A.; Zhang, K.Y.J. NbX: Machine Learning-Guided Re-Ranking of Nanobody–Antigen Binding Poses. Pharmaceuticals 2021, 14, 968. https://doi.org/10.3390/ph14100968

---

Model

The five models inside the model folder are the XGBoost models shown on Figure 1 of our NbX paper.

(i.e. the 5-fold validated models trained and tested on Nb-Ag complexes where pairwise Ag structural alignment quality score < 0.9 to minimize train-test information leakage).

Model	PR-AUC
model_0001	0.229
model_0002	0.169
model_0003	0.349 (best)
model_0004	0.205
model_0005	0.276

---

How to run NbX

To run an example: do step1, step2 then go directly to step5.

To run with your Nb-Ag complex structures: start from step1.

Step 1 : clone and create environment

• Firstly, do...

  git clone https://github.com/johnnytam100/NbX.git

  cd NbX

  conda env create -f environment.yml

  conda activate nbx

  which should have the following libraries installed

  pip install biopandas==0.4.1

  pip install xgboost==0.90

  pip install scikit-learn==0.22.2.post1

  pip install joblib==1.1.0

  pip install dill==0.3.5.1

• Secondly, in the activated nbx environment, manually install...

  (1) PyRosetta (https://www.pyrosetta.org/downloads)

  • NbX was tested on PyRosetta installation with pyrosetta-2022.23+release.f1e0f6d7bf7-cp38-cp38-linux_x86_64.whl from here

  (2) FoldX (https://foldxsuite.crg.eu/)

  (3) DockQ (optional, https://github.com/bjornwallner/DockQ)

  (4) Rosetta (optional, https://new.rosettacommons.org/demos/latest/tutorials/install_build/install_build)

Step 2 : change paths

Change the following paths inside NbX_feature_prep.py

path_to_python = "/home/cltam/anaconda3/envs/nbx/bin/python"

• Get this path inside the activated nbx environment with the command which python

path_to_foldx = "/data/cltam/script/FoldX/foldx_20221231"

path_to_dockq = "/data/cltam/script/DockQ/" (optional, be careful not to omit the last / in this path)

Step 3 : renumber nanobody

Before any docking, please renumber your nanobody with PyIgClassify (http://dunbrack2.fccc.edu/pyigclassify/)

If your Nb (or Nb-Ag complex) structure is confidential and you don't want to submit to a webserver:

modify the CDRs start and end residue numbers (search "CDR1_start_residue") inside NbX_feature_prep.py.

Step 4 : copy your Nb-Ag complex structures to run_NbX folder

cp (path to your Nb-Ag complex structures .pdb) ./run_NbX

Step 5 : change directory

cd run_NbX

cp -r ../model ../NbX_feature_prep.py ../aaDescriptors.csv ../NbX_predict.py ./

Step 6 : feature preparation

(option 1, without DockQ)

python NbX_feature_prep.py --antigen_chain A --antibody_chain H

(option 2, with DockQ)

python NbX_feature_prep.py --antigen_chain A --antibody_chain H --native 6oq8_complex.pdb

Step 7 : predict

python NbX_predict.py

Step 8 : analyze results

Important concept in NbX re-ranking:

Descendingly sort the mean_predicted_CAPRI_binary_proba in NbX_prediction.csv, we get the following results

RosettaDock (optional)

To mimic the NbX benchmark setting, you can perform RosettaDock refinement of your Nb-Ag complex structures before feature preparation.

step 1 : change ROSETTA_PATH in RosettaDock.sh to your Rosetta path.

step 2 : sh RosettaDock.sh

---

Limitations of NbX

1. "Garbage in, garbage out". NbX is not a docking but a re-ranking method, which completely depends on the quality of the input Nb-Ag complex structures to suggest native-like solutions.

   Take action:

   • Use a docking algorithm that is well-tested on predicting native-like Nb-Ab complex structures, no matter how the docking method ranks them.

   • We used Nb-Ag complex structures from ClusPro -> RosettaDock full-atom refinement to benchmark NbX. Please use equivalent or better docking methods.

2. NbX is largely unable to model a single classification threshold that generally applies to all tested Nb-Ag complexes to distinguish non-native-like (0) or native-like (1) Nb-Ag complex structures.

   Take action:

   • Descendingly sort the mean_predicted_CAPRI_binary_proba in NbX_prediction.csv (i.e. the mean native-like probablilty of the 5-fold validated NbX models). This is our NbX re-rank for you.

   • Do consider top ranks as more probable native-like Nb-Ag complex structures compared to the lower ranks.

   • Avoid applying a single classification threshold on the absolute value of the probability to distinguish non-native-like (0) or native-like (1) among unrelated Nb-Ag pairs.

   • Avoid comparing the absolute value of the probability predicted among unrelated Nb-Ag pairs.

   • 202206 Update: Use the following distributions of mean_predicted_CAPRI_binary_proba among 1) crystal 2) native-like and 3) non-native-like Nb-Ag complex structures to guide your selection of native-like Nb-Ag complex structures.