Pose Generation Effects

Paul Francoeur, David Ryan Koes

Department of Computational and Systems Biology, University of Pittsburgh

http://github.com/gnina/

Abstract

Virtual Screening is essential in the drug discovery process, as it reduces all of

chemical space (~10

60

) down to a reasonable number of testable compounds (~10

3

).

Our previous work, gnina, utilized convolutional neural networks to score

protein-ligand binding poses in order to determine if a ligand would bind the

protein. As protein-ligand binding affinity is dependent on its pose, we reason that

there could be benefit to joint training on scoring the protein-ligand pose and

predicting the binding affinity of that pose. We present here an extension to gnina,

which simultaneously predicts a score for the protein-ligand complex and the

affinity of said complex. Additionally we show the importance of training on

docked poses, and testing on clustered cross-validated splits of the training data in

order to obtain a model whose predictions are pose sensitive and generalizable to

unseen data, and showing the importance of proper training data.

Refined

PDBbind 2016 refined set

4057 complexes

69,780 ligand poses

Complete affinity data

Redocked

Subset of Cross-Docked

2923 distinct pockets

790,954 ligand poses

Affinity data for ~40%

Datasets

Smina docked and minimized poses are used for training.

Cross-Docked

Structures from Pocketome

2923 distinct pockets

22,767,152 non-redundant ligand poses

Affinity data for ~40% of ligands

Pose Sensitivity

Predicting Affinity Performance

Models

Def2017

Def2018

Training Protocol

Data Representation

24x24x24Å grid at 0.5Å resolution

14 ligand and 14 receptor atom types

Continuous Gaussian density

CUDA optimized grid generation

Background

Importance of Good Training Data

Protein-Ligand Binding Affinity Prediction with GNINA

Protein-ligand scoring provides a metric of binding strength

between small molecules and target proteins; a critical

subroutine of structure-based drug design. An ideal scoring

function would correctly predict the binding affinity and

correctly identify an accurate ligand pose for the protein.

Convolutional neural networks are state-of-the-art in image recognition.

Convolutional layers apply a small non-linear kernel function iteratively across the

input to produce a feature map. More convolutions are then applied to these feature

maps to recognize higher order features in the input.

Data augmentation is performed by

applying random rotations and

translations (±6Å) to protein-ligand

complex structures. This reduces

overfitting and compensates for the

coordinate-frame dependency of a

3D grid representation.

To extend our previous models, we now

perform joint training on the pose of a

complex with a logistic loss (classification)

AND a mean squared error L2 loss for

affinity prediction (regression). Notably, we

only penalize poor poses for over

predicting the affinity of the complex.

Training upon PDBbind refined-core and testing on the core set, like previous

attempts at this task, yields overly optimistic results. A better measure of

generalizability would be to utilize cross-validated sets for training and testing.

Acknowledgements

This research was supported by R01GM108340 from the National Institute of General Medical Sciences and

contributions from aigrant.org, Google Cloud, NVIDIA Corporation, the University of Pittsburgh Center for

Simulation and Modeling, and the University of Pittsburgh Center for Research Computing.

We observe that in general

there is a left shift (IE more

negative correlations) when

joint training with the Pose

and Affinity as expected.

Observe the inconsistent

performance drop when

crystal poses are removed

from the test set. It is

unclear if this is due to the

model detecting differences

between crystal and docked

poses, or simply a lack of

positive examples when

training. The lack of drop

in affinity prediction

suggests pose information

is not being utilized.