Better AI Molecular Models Need Many Short Simulations, Not One Long Run

Artificial intelligence and machine learning models trained on molecular dynamics trajectories are increasingly being used to predict properties reported from molecular dynamics simulations, spanning chemical, biological, and materials systems. But according to a commentary published in BMC Artificial Intelligence by Alya A. Arabi and Anju Choorakottayil Pushkaran, the reliability of these models depends critically on the quality and reproducibility of the underlying simulations, and a frequent assumption in the field—that longer single simulations necessarily yield more accurate results—may often be wrong. Unless mechanistic insights are the primary goal, the authors argue, multiple shorter, independent simulations are likely to provide a more representative sampling of highly complex potential energy surfaces, which are often characterized by numerous conformational states and local minima.

The stakes are considerable. Molecular dynamics simulations generate extensive time-series data, including conformational changes, binding events, geometrical parameters, hydrogen bonds, secondary structure evolution, domain motions, and free energies. AI workflows can use this data in training, and molecular dynamics data is now commonly used to train artificial intelligence models to rapidly predict outcomes that would otherwise require significantly more time and computational resources to obtain. One example cited in the commentary is UnbindingFlow, a deep generative model trained on an extensive dataset of roughly 13 million frames of dissociation molecular dynamics trajectories, which was used to predict dissociation pathways. Another model, MDtrajNet-1, combines equivariant neural networks with Transformer models to predict molecular dynamics trajectories across diverse chemical systems. A separate Generative Pre-trained Transformer approach discretizes trajectories into different physicochemical states to predict future conformational transitions and kinetic pathways.

Just like any other AI model, however, the quality and accuracy of such predictions greatly depend on the quality of the data used to train the model. For models built using data from molecular dynamics simulations, a single trajectory provides limited sampling and may not capture the full complexity of a given system. Reproducibility in molecular dynamics is a known challenge, since outcomes depend on the input parameters. Each run explores a single trajectory, which may become trapped in a local minimum. Using data from such a single trajectory to train an AI model can lead to biased learning, because the model is provided with only a smaller subset of the conformational and energetic states of the system. Without multiple, statistically independent trajectories, the model may learn artifacts specific to that particular trajectory, such as transient conformations or local energy minima, and biased or poorly sampled training data can propagate errors and reduce predictive power.

The case for replicas rests on a substantial body of prior work. Many articles by several research groups, as well as a recent guideline published by the Journal of Chemical Information and Modelling, highlight the importance of ensemble-based molecular dynamics simulations. These studies clearly show that estimates obtained from individual simulation replicas vary considerably, so increasing the number of independent replicas provides a more diverse and representative dataset for training AI and machine learning models. To make replicas genuinely independent, each simulation can be started with a different geometry, although starting each simulation with different random seeds for velocities is also possible in practice. Varying the velocities or starting geometries effectively mitigates pseudoreplication across the dataset, ensures statistically sound foundations, and enables independent exploration of different regions of the potential energy surface by each trajectory.

The importance of replica simulations is particularly evident in free energy calculations, which can be estimated using different approaches, such as direct sampling of bound and unbound states using alchemical transformations—free energy perturbation or thermodynamic integration—or end-point methods like Molecular Mechanics/Poisson–Boltzmann Surface Area and Molecular Mechanics/Generalized Born Surface Area. Coveney and colleagues quantified the errors in energy averages obtained from ensemble simulations, showing via bootstrap analyses how errors in predicted binding energies decreased with the number of replicas. They demonstrated that using 25 replicas of at least 4 nanoseconds each is necessary to reliably predict and rank binding free energies in protein-ligand systems. In related work, Knapp and colleagues showed that two independent replicas of a given protein produced different mean intra-chain hydrogen bond counts that did not necessarily match experimental data, while the mean count obtained from 100 independent replicas was consistent with the experimental crystal structure.

That study, along with one by Day and colleagues, concluded that at least five to ten replicas are required to reliably estimate average properties in protein systems, with the required simulation length depending on the purpose of the study. When investigating large conformational changes in proteins, protein-protein association, or folding processes, both the number of replicas and the simulation lengths are equally important. Similar conclusions have emerged from nucleic acid systems. Pushkaran and Arabi showed that binding energies of DNA-intercalator systems can deviate by as much as 10 kcal/mol among different replicas; after bootstrap analyses, they proposed that at least eight replicas of 10 nanoseconds each are required to reproduce experimental binding energies in these systems.

Nucleic acid work also illustrates how aggregated shorter runs can match one very long trajectory. Galindo-Murillo and colleagues showed that in a DNA system, independent shorter simulations of roughly 1 microsecond each, when aggregated, reproduced the results of an extremely long single simulation of 44 microseconds. Yan and colleagues showed that 10 independent simulations are required to reproduce the potential energy distribution of RNA with minimal standard errors. Analyzing the folding of an RNA aptamer, they found that different long simulations initiated from different RNA structures got trapped in different local conformations, whereas an ensemble of 60 replicas—10 replicas across six different initial geometries—revealed the full range of aptamer states and transitions. This highlights that considering multiple starting geometries is also important, particularly when investigating mechanisms.

For researchers incorporating multiple simulations into AI workflows, the commentary offers concrete planning guidance. It is best to have at least three replicas to allow basic uncertainty estimation in the values fed to machine learning models, though five to ten replicas are favorable for more reliable statistics and better exploration of the potential energy surface. There is no one-size-fits-all length: chemical or physical intuition, or prior studies, should determine a reasonable length for each replica depending on the objective of the model. Free energy calculations often converge within a few nanoseconds, whereas processes involving major conformational changes usually require significantly longer simulations. As an illustration, running five independent 20-nanosecond trajectories often yields better training data than running a single 100-nanosecond trajectory of the same system. If the standard deviation across replicas is high, the authors advise extending the duration of individual simulations or increasing the number of replicas until that standard deviation stabilizes.

Ensemble-trained AI is not merely theoretical. The AI framework Dynaformer was trained on data from ensemble simulations of various protein-ligand complexes to predict binding affinities, and it outperformed traditional scoring models trained on single static crystal structures. In another study, De Fabritiis and colleagues evaluated a hybrid neural network potential/molecular mechanics model developed for predicting relative binding free energies, with training completed using an ensemble of three replicas per ligand pair. From a workflow perspective, the commentary recommends automating ensemble generation with reproducible pipelines, consistently documenting input parameters such as velocity random seeds and initial geometry, and storing the data associated with each trajectory, since these details are important to ensure the dataset is diverse and thereby improve the generalizability of the resulting model. When gathering information from multiple replicas, statistical processing to produce weighted averages is usually preferable, and combining ensemble data with dimensionality reduction techniques such as clustering or principal component analysis can help identify dominant motions and representative structures.

The broader message is that as AI-guided molecular modeling grows, maintaining diversity in molecular dynamics training data becomes a scientific necessity rather than a technical nicety. Each additional replica contributes not only to reducing uncertainty but also to strengthening the scientific foundation and generalizability of AI predictions. The commentary is published as an open-access article in BMC Artificial Intelligence, and no datasets were generated or analysed during the study. For a field increasingly reliant on simulation-derived training data, the authors’ central recommendation is straightforward: researchers who adhere to these ensemble practices can build AI models that are both accurate and reliable, while those who rely on single, poorly sampled trajectories risk embedding hidden artifacts directly into their models’ predictions.

Subject of Research: Best practices for training AI models on molecular dynamics simulation data

Article Title: Artificial intelligence models trained on data from molecular dynamics trajectories

Article References: Arabi, A. A., & Pushkaran, A. C. (2026). Artificial intelligence models trained on data from molecular dynamics trajectories. BMC Artificial Intelligence, 2(1), Article 3. https://doi.org/10.1186/s44398-026-00021-0

Image Credits: AI Generated

DOI: 10.1186/s44398-026-00021-0

Keywords: artificial intelligence, machine learning, molecular dynamics, ensemble simulations, free energy calculations, protein-ligand binding, training data quality, reproducibility, potential energy surface, nucleic acids, sampling, computational chemistry