Documentation

pHaseMD4AI is a public-facing data portal for the exploration of molecular dynamics datasets spanning peptides and proteins under chemical modification and pH-dependent perturbation conditions. The website is designed as a centralized access point for browsing curated simulation entries, inspecting residue and modification definitions, reviewing workflow documentation, and obtaining data for downstream computational analysis.

The current version of the portal integrates peptide systems, protein systems, post-translationally modified variants, and constant-pH molecular dynamics cases within a unified interface. Users can search records by sequence, protein identifier, modification label, or protonation-state notation, and can navigate from summary tables to structure previews and entry-specific detail pages. This organization is intended to support both rapid dataset discovery and more targeted investigation of related molecular systems.

In addition to dataset access, the website documents the simulation and analysis framework used to generate and organize the data. The workflow description covers conventional unbiased molecular dynamics, adaptive sampling, conformational-state analysis, kinetic modeling, and downstream interpretation of free-energy landscapes and transition behavior. By combining data presentation with methodological context, the portal helps users understand not only what data are available, but also how the data were produced and how they may be reused.

The website also includes residue-centric reference pages for standard amino acids, modified residues, phosphorylation states, and protonation variants, as well as statistical summaries describing dataset composition and scale. Together, these components make pHaseMD4AI both a database interface and a documentation resource for molecular simulation studies involving sequence variation, chemical modification, and pH-dependent conformational behavior.

The workflow of our main simulation strategy is designed to identify the main conformational states of a system, describe the kinetic relationships between them, and determine which regions of conformational space still require additional sampling.

The procedure has two connected parts: sampling and kinetic analysis. The analysis results are then used to guide further sampling.

The sampling stage combines two approaches: conventional unbiased molecular dynamics simulation and adaptive sampling.

The unbiased simulation provides a dynamical baseline. Starting from a given initial conformation, the system evolves continuously without additional bias and produces a single long trajectory. In the current setup, the total length of this baseline trajectory is 1 microsecond.

Adaptive sampling divides the total sampling effort into multiple rounds of parallel short trajectories. In the current setup, each round consists of 10 parallel trajectories, each 10 ns in length, and a total of 5 rounds are carried out, corresponding to an overall sampling time of 500 ns.

After each round, structural features are extracted from the accumulated trajectories and used to evaluate which regions of conformational space have already been explored and which remain under-sampled.

The core features used for adaptive sampling are geometric descriptors, mainly pairwise distances between atoms that reflect global conformational variation.

Based on these features, low-dimensional representations, clustering results, and kinetic models are constructed. Conformations expected to provide the most useful new information are then selected as starting points for the next round.

Once sampling is complete, all trajectories are analyzed within a unified framework. The purpose of the analysis is to organize the raw trajectory data into a state-space description with kinetic meaning.

The analysis begins with structural feature extraction. The same feature definition used during adaptive sampling is also used here, so that sampling and analysis are based on a consistent representation of conformational space.

After high-dimensional structural features are obtained, dimensionality reduction is performed to extract low-dimensional variables that capture the dominant slow conformational motions of the system.

Conformations are then clustered in this low-dimensional space, which discretizes the conformational landscape into a set of microstates. On the basis of these microstates, a Markov state model is constructed.

The microstates are further coarse-grained into a smaller number of macrostates with clearer physical interpretation. In parallel, the model is subjected to timescale and kinetic consistency checks.

This analysis yields the free-energy landscape, mean first-passage times between states, dominant transition pathways, and representative structures for the major states.

The final outcome is not a collection of isolated trajectories, but a kinetic state-space description that can be used directly to interpret the conformational behavior of the system.

This framework is intended to answer several central questions: what the major conformational states are, how these states interconvert, which states are more stable, which transitions occur on slower timescales, which regions have already been explored, and which regions remain worth additional sampling.

In practice, the workflow provides representative structures, free-energy landscapes, kinetic connectivity, mean transition times, differences in conformational coverage, and seed conformations for later simulations or downstream analysis.

Reuse of this workflow can be organized directly around four core scripts.

If the goal is to run a conventional long baseline simulation, refer to mdrun.sh.

If the goal is to carry out multi-round adaptive sampling, with the analysis from each round automatically determining the starting conformations for the next round, refer to AS.py.

If the goal is to analyze a single baseline trajectory, refer to analysis_unbiased.py.

If the goal is to analyze the complete set of adaptive sampling trajectories and compare adaptive sampling against the baseline simulation, refer to analysis_AS.py.