RosENet: Improving binding affinity prediction by leveraging molecular mechanics energies with a 3D Convolutional Neural Network

Hussein Hassan-Harrirou, Ce Zhang, Thomas LemminbioRxiv 2020 · DOI

Abstract

The worldwide increase and proliferation of drug resistant microbes, coupled with the lag in new drug development represents a major threat to human health. In order to reduce the time and cost for exploring the chemical search space, drug discovery increasingly relies on computational biology approaches. One key step in these approaches is the need for the rapid and accurate prediction of the binding affinity for potential leads. Here, we present RosENet (Rosetta Energy Neural Network), a three-dimensional (3D) Convolutional Neural Network (CNN), which combines voxelized molecular mechanics energies and molecular descriptors for predicting the absolute binding affinity of protein–ligand complexes. By leveraging the physico-chemical properties captured by the molecular force field, our model achieved a Root Mean Square Error (RMSE) of 1.26 on the PDBBind v2016 core set. Our study demonstrated that molecular mechanics energies can be voxelized and used to help improve the predictive power of the CNNs.

Introduction

In the race to discover new drugs, scientists face a staggering chemical search space estimated at $10^{66}$ molecules. Because experimental testing of every candidate is impossible, drug discovery relies heavily on computational scoring to identify potential leads. However, current methods face a significant trade-off: physics-based simulations (like Free Energy Perturbation) are highly accurate but computationally expensive, while standard Machine Learning (ML) models are fast but often rely on simplified descriptors that miss nuanced physical interactions like solvation effects or electrostatic forces.

RosENet (Rosetta Energy Neural Network) bridges this gap by feeding raw physical energy data directly into a deep learning architecture. By treating a protein-ligand complex like a 3D image and "voxelizing" physics-based energies, the model captures the complex spatial relationships that drive molecular recognition.

Methodology

The RosENet pipeline transforms a protein-ligand complex into a multi-channel 3D grid suitable for convolutional neural networks.

The RosENet workflow: from dataset selection and Rosetta minimization to feature embedding and affinity prediction.

Voxelization and Feature Engineering

The binding pocket is mapped into a $25 \times 25 \times 25$ Å grid of voxels (3D pixels). Each voxel contains 20 channels of information:

  • Molecular Descriptors (4 channels): Identifying aromatic carbons, H-bond acceptors, and ionizable groups.
  • Molecular Mechanics (MM) Energies (6 channels): Derived from the Rosetta force field, these include attractive/repulsive forces, electrostatics, and implicit solvation energies.
Visualization of the voxel grid and the spatial distribution of different energy channels (Attractive, Electrostatic, and Implicit Solvation) within the binding pocket.

Model Architecture and Training

The researchers utilized a modified ResNet-101 architecture. They hypothesized that ResNet’s "skip connections" are ideal for learning the subtle variations in energy that determine binding affinity. Before voxelization, every complex undergoes a "relaxation" step using Rosetta to ensure the ligand is in a physically plausible, low-energy pose.

The model was trained and validated using the PDBBind dataset (v2016 and v2018). To ensure robustness, it was tested on several independent benchmarks, including CSAR (NRC-HiQ), the Astex Diverse Set, FEP datasets, and PDBBind NMR structures.

Results and Discussion

RosENet achieved a Root Mean Square Error (RMSE) of 1.26 on the PDBBind v2016 core set, outperforming popular CNN-based models like OnionNet and Pafnucy on several benchmarks.

Physics-Enhanced Generalization

A key insight from the study is that while models using only molecular descriptors perform well on familiar data, adding MM energies helps the model generalize much better to entirely new protein families. Specifically, energy terms like implicit solvation provide critical information about how water molecules are displaced during binding—information that simple atom-type descriptors usually lack.

Architecture and Data Bias

The study found that deep architectures are necessary for this task; a shallower SqueezeNet model showed significantly lower performance, suggesting that the relationship between 3D physical energies and binding affinity is highly non-linear. However, like many ML models, RosENet is most accurate in the "medium" affinity range ($pK_D$ 6 to 8), highlighting a persistent challenge in the field: the need for more high-quality data on very weak and very strong binders.

Conclusion

RosENet demonstrates that molecular mechanics energies can be effectively voxelized to improve the predictive power of 3D CNNs in drug discovery. Future work will focus on incorporating features from Molecular Dynamics (MD) simulations to capture protein flexibility and exploring ensemble methods to further reduce prediction error.

Summary updated

Abstract

ABSTRACT The worldwide increase and proliferation of drug resistant microbes, coupled with the lag in new drug development represents a major threat to human health. In order to reduce the time and cost for exploring the chemical search space, drug discovery increasingly relies on computational biology approaches. One key step in these approaches is the need for the rapid and accurate prediction of the binding affinity for potential leads. Here, we present RosENet ( Ros etta E nergy Neural Net work), a three-dimensional (3D) Convolutional Neural Network (CNN), which combines voxelized molecular mechanics energies and molecular descriptors for predicting the absolute binding affinity of protein – ligand complexes. By leveraging the physico-chemical properties captured by the molecular force field, our model achieved a Root Mean Square Error (RMSE) of 1.26 on the PDBBind v2016 core set . We also explored some limitations and the robustness of the PDBBind dataset and our approach, on nearly 500 structures, including structures determined by Nuclear Magnetic Resonance and virtual screening experiments. Our study demonstrated that molecular mechanics energies can be voxelized and used to help improve the predictive power of the CNNs. In the future, our framework can be extended to features extracted from other biophysical and biochemical models, such as molecular dynamics simulations. Availability https://github.com/DS3Lab/RosENet
BibTeX
@misc{hassanharrirou2020rosenet,
  author = {Hussein Hassan-Harrirou and Ce Zhang and Thomas Lemmin},
  title = {{RosENet: Improving binding affinity prediction by leveraging molecular mechanics energies with a 3D Convolutional Neural Network}},
  year = {2020},
  doi = {10.1101/2020.05.12.090191},
  howpublished = {bioRxiv},
}
Figure