Condensation of Force Field Parameters from Machine Learning Predicted Distributions for High-Throughput Virtual Screening Applications

Domenico Bonanni, Yuedong Zhang, Davide Gadioli, Gianluca Scarpellini, Pietro Morerio, Alessio Del Bue, Andrea Rosario Beccari, Gianluca PalermoJournal of Chemical Information and Modeling 2025 · DOI

TL;DR

This paper introduces a novel "condensation" method to drastically speed up machine learning force fields (MLFFs) for high-throughput virtual screening (HTVS). By statistically summarizing MLFF-predicted force field parameters into a lookup table, they achieve a 30x speedup with only a minor loss in accuracy, making MLFFs practical for large-scale screening.

What Problem Does This Paper Address?

Developing accurate molecular force fields is crucial for simulating molecular behavior, but traditional methods are computationally expensive. Machine learning force fields (MLFFs) offer high accuracy and flexibility, but their runtime prediction times are too slow for high-throughput virtual screening (HTVS), a process that requires evaluating millions of molecules. Existing MLFFs, like Espaloma, generate parameters on-the-fly, creating a bottleneck. The challenge is to retain the accuracy of MLFFs while making them fast enough for HTVS.

Methods

The core innovation is a "condensation" approach. Instead of predicting force field parameters for each molecule at runtime, the authors first use the MLFF (Espaloma) to generate a distribution of parameters for a large set of molecules. This distribution, which captures the variability of parameters across different chemical environments, is then condensed into a single representative value (e.g., the mean) for each parameter type. These condensed values are stored in a lookup table, allowing for rapid retrieval during HTVS. This bypasses the computationally intensive MLFF inference step. They also explored using both Class I and Class II functional forms for the force fields.

Datasets

  • Training/Parameter Generation: The authors used five subsets from the OpenFF QCArchive: SPICE-PubChem, SPICE-DES-Monomers, SPICE-Dipeptide, Gen2-Opt, and Gen2-Torsion, totaling 17,794 molecular structures.
  • Evaluation: The OpenFF Industry Benchmark Season 1 v1.1 dataset was used to evaluate the predictive accuracy of the force fields. This dataset consists of drug-like molecules and their QM-optimized geometries, serving as ground truth.

Key Insights and Ideas

  • Statistical Condensation is Key: The core idea that a statistical summary (like the mean) of MLFF-predicted parameter distributions can represent the parameter's behavior across diverse chemical environments is a significant conceptual leap. It bridges the gap between continuous ML predictions and discrete force field parameters.
  • Speed vs. Accuracy Trade-off is Manageable: The paper demonstrates that a 30x speedup can be achieved with only a minor decrease in accuracy, suggesting that for HTVS, this trade-off is highly favorable.
  • MLFFs Can Be Made HTVS-Compatible: This work provides a concrete pathway to adapt powerful MLFFs for applications where computational speed is paramount, opening up new possibilities for drug discovery and materials science.
  • Class II Limitations: The study notes that while Class II functional forms are theoretically more accurate, training MLFFs with them presents significant optimization challenges due to ill-conditioned loss functions.

Key Results

  • Computational Speedup: The condensed Espaloma force field achieved a 30-fold speedup compared to the original Espaloma approach.
  • Accuracy: The condensed force field showed only a minor decrease in predictive accuracy, with RMSD and TFD values exhibiting a slight increase compared to the original Espaloma model. For instance, the density difference for RMSD was within 0.1.
  • Performance Comparison: The condensed force field performed comparably to, and in some metrics better than, established transferable force fields like MMFF94, while being significantly faster than the original Espaloma Class I and Class II models.
  • Table 1 Summary: The table shows that the condensed Espaloma model achieved competitive RMSD and TFD values, often outperforming MMFF94 and Espaloma Class I, while being vastly faster.

Limitations and Future Work

  • Bonded Parameters Only: The current condensation approach primarily focuses on bonded parameters, with non-bonded parameters borrowed from OpenFF 2.0.0.
  • Potential for Missing Interactions: If a specific atom type combination is not present in the condensation table, that interaction's contribution is disregarded, which could lead to inaccuracies.
  • Future Work:
    • Expand the condensation to include non-bonded parameters.
  • Incorporate more advanced ML techniques like Active Learning to efficiently expand the model's applicability domain.
  • Develop consistent protein-ligand force fields using the same condensation scheme.
  • Adapt the condensation workflow to new atom types or chemical sensing models without extensive retraining.
  • Explore including more complex intermolecular interactions (e.g., H-bonding, pi-pi stacking) for improved protein-ligand binding accuracy.

Summary updated

BibTeX
@article{bonanni2025condensation,
  author = {Domenico Bonanni and Yuedong Zhang and Davide Gadioli and Gianluca Scarpellini and Pietro Morerio and Alessio Del Bue and Andrea Rosario Beccari and Gianluca Palermo},
  title = {{Condensation of Force Field Parameters from Machine Learning Predicted Distributions for High-Throughput Virtual Screening Applications}},
  journal = {Journal of Chemical Information and Modeling},
  year = {2025},
  doi = {10.1021/acs.jcim.5c02199},
}
Figure