Advancing Drug Discovery with AceForce 1.0

We are introducing AceForce 1.0, a neural network potential (NNP) designed to enhance the prediction of molecular interactions. With quantum-level accuracy, AceForce 1.0 addresses key challenges in drug discovery. These include accurately predicting molecular interactions, ranking potential drug candidates, and reducing computational costs. Together, these capabilities offer scientists a robust tool for early-stage therapeutic evaluation.

What Are Neural Network Potentials (NNPs)?

Neural Network Potentials (NNPs) are a class of machine learning models that predict molecular interactions by mimicking quantum mechanical energy surfaces. Unlike traditional force fields, which rely on fixed mathematical equations and often lead to lower accuracy or higher sources of error, NNPs leverage large datasets of quantum mechanical calculations to learn the complex relationships governing atomic forces and energies. In many cases, traditional force fields can be the primary source of error in physics-based methods like binding free energy calculations. NNPs can generalize beyond the limitations of traditional force fields. While traditional methods are often parameterized for specific molecules or environments, NNPs leverage their training on large, diverse datasets to model interactions across a broader chemical space, including rare or highly complex molecular systems. This capability reduces the risk of significant inaccuracies when dealing with atypical chemical structures or conditions.

In the context of drug discovery, NNPs like AceForce 1.0 provide:

• Higher Accuracy: Capturing quantum mechanical interactions leads to more precise predictions in comparison to traditional force fields.
• Improved Generalization: Ability to model a wider range of chemical systems compared to traditional methods.
• Cost Efficiency: By offering quantum-like accuracy at a fraction of the computational cost of direct quantum mechanical calculations.

Features and Capabilities of AceForce 1.0

AceForce 1.0 builds on advanced deep learning methodologies to capture quantum mechanical interactions, which are crucial for understanding atomic forces and energies. Key features include:

• High Accuracy: Trained on a proprietary dataset of millions of quantum mechanical (QM) calculations, AceForce 1.0 mirrors high-level QM methods for reliable molecular interaction predictions.
• Wide Applicability: Its compatibility with diverse chemical elements and charged molecules broadens its utility in drug discovery.
• Efficient Simulations: AceForce 1.0 achieves twice the computational speed of earlier-generation NNPs while maintaining precision.

Performance Validation with QuantumBind-RBFE

To validate its performance, AceForce 1.0 was benchmarked using Acellera’s QuantumBind-RBFE platform against recognized datasets. Relative binding free energy (RBFE) calculations—critical for predicting drug-protein interactions—demonstrated AceForce 1.0's capability to match or exceed the accuracy of leading force fields. This performance was consistent across diverse systems, ensuring reliable ranking of drug candidates. Extended or further information can be found in our recent preprint, where we showcase extensively the capabilities of AceForce 1.0 in RBFE calculations. In addition to RBFE calculations, AceForce 1.0 was evaluated on torsion scan benchmarks, where it demonstrated high accuracy in modeling torsional energy profiles. These tests highlight AceForce's ability to closely replicate quantum mechanical energy surfaces for complex conformational changes, outperforming traditional NNPs like ANI-2x.

‍

Applications in Computational Drug Discovery

AceForce 1.0 integrates seamlessly into computational workflows, offering actionable insights for:

• Screening and prioritizing drug candidates.
• Modeling complex protein-ligand interactions.
• Reducing the computational cost associated with high-accuracy simulations.

Usage

AceForce 1.0 is designed for use in NNP/MM workflows, where the ligand is treated with the neural network potential and the environment with molecular mechanics. It can also be used to simulate pure NNP of small molecules.

• Run ML potential molecular simulations of a small molecule using ACEMD with this tutorial, e.g. to minimize.
• For a tutorial on running mixed protein-ligand simulations, refer to NNP/MM in ACEMD.
• Execute RBFE calculations with QuantumBind-RBFE in an NNP/MM setting with the following tutorial. Further examples can be found here.

Future Developments

Acellera is committed to advancing AceForce’s capabilities. Plans include expanding the training dataset, enhancing computational efficiency, and enabling simulations with larger timesteps to further optimize the balance between cost and accuracy.

Availability

AceForce 1.0 is available for non-profit use and demonstrations, allowing researchers to explore its potential in their projects. The AceForce 1.0 model can be found in huggingface.

‍

This blog post is licensed under a Creative Commons Attribution-ShareAlike 4.0 International License, enabling others to share and adapt the material for any purpose, even commercially, as long as appropriate credit is given and any modifications are distributed under the same license.

‍