AdaptivePy is a compact, extensible Python framework developed to unify implementations of seed-selection policies used in adaptive sampling for molecular dynamics (MD). Adaptive sampling accelerates conformational exploration in MD by iteratively analyzing accumulated trajectories and initiating new simulations from configurations identified as informative. The authors present AdaptivePy to address fragmentation across separate, often incompatible implementations of adaptive sampling methods, which complicates systematic comparisons and combined use in meta adaptive sampling schemes.
The framework implements nine seed-selection policies behind a single configuration-driven interface. These policies span a spectrum from simple population-based baselines to machine-learning and geometry-based methods, and include two ensemble or meta sampling policies introduced in this work. The unified interface is intended as a common testbed for adoption, benchmarking, and continued development of adaptive sampling methods for biomolecular MD simulations.
AdaptivePy exposes nine distinct policies through a consistent configuration-driven API. The collection covers several classes of existing methods: simple population-based baselines; established machine-learning approaches; geometry-based selection methods; and newly introduced ensemble/meta policies. By standardizing how policies are invoked and how seed selections are returned, the framework facilitates direct comparisons between policies and enables users to combine methods without rewriting or adapting disparate codebases.
The authors report that the shared implementation reproduces the characteristic selection behavior of each policy when applied to benchmark problems. That reproducibility across policies under the same interface supports the framework's goal of enabling systematic evaluation and combined use of different strategies for selecting informative starting configurations during adaptive sampling.
A new adaptive sampling scheme is introduced that integrates TS-DAR, a deep learning framework originally designed to identify transition states, as an acquisition criterion. In this work TS-DAR is repurposed so that its outputs drive the selection of seeds with the explicit goal of discovering an entire multi-basin landscape starting from a single basin.
The manuscript describes how TS-DAR's transition-state–focused predictions can be leveraged as an acquisition function within an adaptive sampling loop. Using this approach, the framework aims to prioritize configurations likely to lead to transitions, thereby facilitating exploration of additional basins in the conformational landscape from limited initial conditions.
The authors evaluated AdaptivePy on a series of analytic benchmark landscapes to verify that the implementations reproduce the expected selection patterns of each policy. On those benchmarks, the shared implementation reproduced the characteristic selection behavior of the nine policies, demonstrating consistent behavior across different algorithmic approaches when invoked through the unified interface.
Concrete numerical results, performance metrics, or detailed benchmark figures are provided in the preprint and its supplementary materials; readers should consult those materials for quantitative comparisons and specific experimental setups. The preprint includes a full PDF and supplementary material linked from the source.
Because all policies are implemented behind the same configuration-driven interface, AdaptivePy enables meta adaptive sampling policies that aggregate rankings produced by several individual policies into a single set of seeds. The authors introduce two ensemble or meta sampling policies in this work and describe how aggregation of rankings can be achieved to produce composite seed-selection strategies.
This capability allows users to combine complementary selection criteria — for example, geometry-based cues with machine-learning assessments of transition likelihood — to form hybrid acquisition strategies that may be more robust across different landscapes or initial conditions.
AdaptivePy is presented as a unified testbed to support adoption, benchmarking, and further development of adaptive sampling methods for biomolecular MD. By reducing implementation heterogeneity, the framework seeks to lower barriers to systematic comparisons between algorithms, to enable development of meta policies that aggregate multiple methods, and to facilitate integration of novel acquisition criteria such as TS-DAR into adaptive sampling workflows.
The preprint format indicates that the work has not yet been peer reviewed. The repository, supplementary materials, and PDF are linked from the source record; details about software distribution, documentation, and licensing are available there. The preprint is distributed under a CC-BY-NC-ND 4.0 International license.
Authors are Hassan Nadeem, Diego E. Kleiman, and Diwakar Shukla, affiliated with the University of Illinois at Urbana-Champaign. Correspondence is directed to Diwakar Shukla at diwakar@illinois.edu as noted in the source. The authors declare no competing interests.
Funding reported for the work includes support from the Army Research Office (W911NF-22-2-0246) and the National Institutes of Health (R35GM142745). The preprint was posted August 06, 2026, and supplementary material and the full PDF are available from the source record.