🧬 biology

MDFitML: Using machine learning to predict potency from molecular simulations

MDFitML introduces an automated workflow that leverages machine learning models trained on simulation fingerprints (SimFP) derived from molecular dynamics trajectories to accurately predict and interpret the potency of diverse ligands by mapping predictions directly to specific residue-level protein-ligand interactions.

Original authors: Xue Zong, Ahmet Mentes, Barmak Mostofian, Keith Jones, David R Stevens, Greg Bakken, Sirish Kaushik Lakkaraju

Published 2026-08-06

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Xue Zong, Ahmet Mentes, Barmak Mostofian, Keith Jones, David R Stevens, Greg Bakken, Sirish Kaushik Lakkaraju

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

Imagine you are trying to find the perfect key to open a very tricky, shape-shifting lock. In the world of medicine, this lock is a protein inside your body, and the key is a new drug molecule. For decades, scientists have tried to design these keys by looking at a single, frozen photograph of the lock. But real locks aren't static; they wiggle, breathe, and change shape. To understand how a key really fits, you need to watch the lock in motion, like watching a video instead of a still photo. This is where "molecular dynamics" comes in: it's a supercomputer simulation that acts like a high-speed movie, showing how proteins and drug molecules dance and interact over time. However, these movies generate mountains of data, making it hard to figure out exactly which tiny movement makes a drug strong or weak. This is the puzzle the researchers tackled: how do we turn these complex, moving movies into a simple recipe that predicts how well a drug will work?

Enter MDFitML, a new automated workflow that acts like a smart translator between these molecular movies and drug design. The researchers combined two powerful tools: the "movie camera" of molecular dynamics simulations and the "pattern-spotting" power of machine learning. Instead of just looking at the final pose of a drug, their method watches the entire simulation to create a unique "fingerprint" called a SimFP (Simulation Fingerprint). Think of this fingerprint as a scorecard that counts every time the drug hugs, high-fives, or holds hands with specific parts of the protein during the simulation. It captures the stability of these interactions, noting which ones are strong and lasting, and which are fleeting and weak.

The team tested this approach on eight different protein targets, including some famous ones like CDK2 and BACE1, using datasets ranging from 16 to 147 different drug-like molecules. They found that by feeding these interaction fingerprints into a machine learning model, they could accurately predict the "potency" (how strong the drug is) of the molecules. In fact, for most of the targets they studied, their new method was much better than the old way of just looking at static snapshots. For example, with the CDK2 protein, their model's ability to explain the differences in drug strength jumped from a score of 0.57 to a very impressive 0.90. With MCL-1, the improvement was even more dramatic, going from a weak 0.10 to a strong 0.70.

What makes this truly special is that the model doesn't just give a number; it tells you why. Because the model learns from the specific interactions seen in the simulation, it can point out exactly which amino acid in the protein is responsible for a drug's success. In one case study involving CDK2, the model identified that a specific interaction with a part of the protein called Lys89 was the secret sauce for strong binding. It showed that drugs with a specific chemical group (a sulfonamide) that could grab onto Lys89 were much stronger than those that couldn't.

Perhaps the most exciting discovery is that this method is surprisingly flexible. Usually, if you train a computer model on one type of drug shape, it fails miserably when you show it a completely different shape. But MDFitML suggested that as long as the new drug fits into the same "pocket" of the protein, the model could predict its strength, even if the drug looked totally different from the ones it was trained on. They tested this by training the model on one group of drugs (nicotinamides) and then asking it to predict the strength of a completely different group (imidazopyridazines) targeting the same protein spot. The model managed to rank-order the new, diverse drugs with reasonable accuracy. This suggests that by focusing on the dance between the drug and the protein, rather than just the drug's chemical structure, scientists might be able to design better medicines even when they are exploring new, unfamiliar chemical territories.

The researchers are careful to note that this isn't a magic wand that solves every problem instantly. Their results are based on simulations and specific datasets, and the method works best when the new drugs occupy the same binding pocket as the training drugs. However, by turning complex, moving molecular data into clear, interpretable rules, MDFitML offers a promising new way to understand why some drugs work and others don't, potentially speeding up the journey from the lab bench to the pharmacy shelf.

Technical Summary: MDFitML – Using Machine Learning to Predict Potency from Molecular Simulations

Problem Statement
While Molecular Dynamics (MD) simulations provide a physics-based, time-resolved description of protein-ligand interactions that static docking poses cannot, their application to quantitative affinity prediction in drug discovery remains limited. Current challenges include the lack of scalable, automated workflows to process large numbers of complexes and the difficulty of reducing high-dimensional trajectory data into descriptors that are both predictive and chemically interpretable. Conversely, standard machine learning (ML) approaches for potency prediction typically rely on static chemical structures or docked poses, failing to explicitly encode the dynamic binding behaviors and conformational flexibility that distinguish closely related analogs. There is a need for a method that systematically transforms MD trajectories into concise, interpretable features suitable for the small, heterogeneous datasets common in early discovery.

Methodology
The authors present MDFitML, an automated workflow that integrates MD simulations with ML models using Simulation Fingerprints (SimFP) as features. The workflow proceeds as follows:

System Preparation & Simulation: Protein-ligand complexes are prepared (using OpenEye Spruce Toolkit) and subjected to all-atom MD simulations (100 ns) using the OpenMM engine, AMBER14SB force field for proteins, and OpenFF 2.0.0 for ligands. Simulations are run in triplicate to increase sampling instances.
Feature Extraction (SimFP): Instead of static interaction counts, the workflow calculates interaction occupancy from the MD trajectories. This metric quantifies the stability of interactions (contact, hydrogen bonds, salt bridges, stacking, and water-mediated interactions) as the percentage of simulation time an interaction is observed. These occupancies serve as the "SimFP" descriptors.
Model Development Strategy: To address the "curse of dimensionality" (many interaction features vs. few compounds), a two-step modeling approach is employed:
- Feature Selection: A Lasso regression model with L1 regularization is applied to the full feature set to identify the most informative interactions and handle multicollinearity.
- Prediction: A standard linear regression model is trained using only the top 10 interactions selected by the Lasso model.
- Validation: Models are validated using nested leave-one-molecule-out cross-validation (LOMO-CV) and Y-scrambling experiments to rule out spurious correlations.
Ensemble Prediction: For final potency predictions, the model averages predictions from the three replicate MD simulations per compound to reduce variance.

Key Contributions

Simulation Fingerprints (SimFP): The introduction of a new descriptor class derived from interaction occupancies over MD trajectories. Unlike static descriptors, SimFP enables interpretable ML models where every prediction maps directly to specific residue-level pocket interactions.
Cross-Chemotype Transferability: The study demonstrates that linear models trained on SimFP from a single chemotype can accurately rank-order structurally diverse compounds targeting the same pocket. This capability extends beyond the training set's chemical space, a feat often challenging for both relative free energy methods and descriptor-based ML.
Automated, Scalable Workflow: The integration of MD setup, trajectory analysis, and ML training into a fully automated pipeline (enabled by the Orion compute infrastructure) allows for the simultaneous prediction of potency and mechanistic interpretation in data-limited settings.

Results
The workflow was validated across eight systems, including five public targets (CDK2, CDK9, MCL-1, BACE1, Hif-2α) and two internal Bristol Myers Squibb (BMS) targets with diverse chemotypes.

Performance vs. Static Poses: MDFitML consistently outperformed models built on static docked poses. For example, the coefficient of determination ( $R^2$ ) for CDK2 improved from 0.57 to 0.90, and for MCL-1 from 0.10 to 0.70. The dynamic features captured transient interactions and solvent effects that correlated better with experimental potencies.
Feature Selection Efficacy: Using the top 10 Lasso-selected interactions yielded superior predictive accuracy compared to full Lasso models or models with more features, effectively balancing sample-to-feature ratios and preventing overfitting.
Mechanistic Interpretation: The models successfully identified specific structural rationales for potency differences. In the CDK2 case study, the model correctly identified that tight binding correlated with specific contacts with Lys89, distinguishing potent ligands from weak binders based on the arrangement of sulfonamide groups.
Cross-Chemotype Prediction:
- Tyk2: A model trained on nicotinamide chemotypes successfully predicted the potency of imidazopyridazine (IZP) compounds, despite significant structural differences.
- BMS Internal Target: A model trained on a biaryl scaffold (Chemotype A) successfully rank-ordered an acetylene scaffold series (Chemotype B).
- This suggests that as long as query molecules occupy the same binding pocket and can be featurized into the SimFP space defined by the training set, the model can generalize across chemical diversity.

Significance and Claims
The paper claims that MDFitML offers a robust alternative to established free energy and pose-based ML methods, particularly in data-limited regimes. Its primary significance lies in interpretability: by learning potency differences via MD interactions, the model provides an unprecedented structural rationale for predictions that is often lost when using high-dimensional chemical descriptors.

The authors emphasize that while both Free Energy Perturbation (FEP) and standard ML models suffer from reduced reliability when query molecules diverge significantly from the training chemical space (due to phase space overlap issues or structural dissimilarity), MDFitML mitigates this by focusing on the pocket interaction space. As demonstrated with Tyk2 and the BMS internal target, the method allows for reliable traversal of large chemical distances while maintaining interpretability, provided the query ligand binds to the same pocket. The workflow is designed to be scalable and reproducible, offering a practical tool for design triaging in drug discovery projects.

Technical Summary: MDFitML – Using Machine Learning to Predict Potency from Molecular Simulations

More like this