Dataset for ProtXAI: Explainable AI Analysis of Protein Dynamics

This dataset accompanies the study “ProtXAI: Explainable AI Reveals Structural Determinants of Protein Dynamics” and contains molecular dynamics (MD) simulation datasets used to develop and evaluate the ProtXAI framework.

ProtXAI is an integrated molecular dynamics, machine learning (ML), and explainable artificial intelligence (XAI) framework designed to identify residue-level structural determinants of protein dynamics. The approach converts MD trajectories into structural and dynamical representations and uses deep learning models together with layer-wise relevance propagation (LRP) to identify residues and residue–residue interactions that contribute to conformational changes, ligand-induced effects, mutation-dependent dynamics, and evolutionary changes.

The dataset contains simulation data for three representative protein systems:

1. ApoE4 [ApoE4.tar]

• Apolipoprotein E4 simulated in the absence and presence of the small-molecule ligand 3-sulfopropanoic acid (3-SPA).

• The simulations were generated using adaptive sampling, with an aggregate simulation time of 20 μs.

• Frames were subsampled every 5 ns and restricted to residues 24–161, resulting in 8,036 MD frames.

• These data were used to train a binary classification model distinguishing the apo and ligand-bound states and to identify residue-pair interactions associated with ligand-induced changes in protein dynamics.

2. Staphylokinase (SAK) variants [SAK.tar]

• MD simulation data for four engineered SAK variants: SAK42D, STAR, SY155, and THR174.

• Each variant was simulated for 200 ns across three independent replicas, corresponding to 600 ns per variant.

• Snapshots were extracted every 0.1 ns, resulting in 23,956 frames in total.

• These data were used for four-class classification and XAI analysis to identify variant-specific dynamic hotspots and residue–residue interactions.

• SAK data numbering: The sample/frame indices in the deposited SAK dataset correspond to the variants as follows:

3. Renilla luciferase subfamily [RLuc.tar]

• MD simulation data for three proteins in the Renilla luciferase evolutionary series: AncHLD-RLuc (PDB ID: 6G75), RLuc8 (PDB ID: 2PSF), and AncFT (PDB ID: 6S97).

• The trajectories were aligned to reference structures and subsampled every 5 ns.

• Cα Cartesian coordinates were extracted for machine-learning analysis.

• The data were used in a self-supervised next-frame prediction framework to investigate intrinsic protein dynamics and the redistribution of conformational motions along the luciferase evolutionary trajectory.

4. PyMOL visualization session [pymolsession_isomesh_ApoE4.pse]

• PyMOL session file associated with Figure 2b of the manuscript.

• The session contains the molecular visualization used to display the ligand-associated density/isomesh on the ApoE4 structure.

• The isosurface/isomesh level can be adjusted in PyMOL using the following command: isomesh isomap, checkpoint_density_out, 0.15. The final value (0.15) represents the isovalue used to generate the displayed mesh. Users can modify this value to change the density threshold and visualization.

For the ApoE4 and SAK systems, the ProtXAI workflow uses residue–residue distance-difference representations to capture short-timescale changes in protein interactions. For the luciferase systems, Cα Cartesian coordinates are used for next-frame prediction.

These datasets support the analyses reported in the accompanying study, including the identification of experimentally supported dynamic hotspots and additional long-range structural couplings using explainable machine learning. The datasets are intended to facilitate reproducibility, benchmarking, and further development of computational approaches for automated analysis and interpretation of protein molecular dynamics trajectories.

Relevance values: The residue-level and residue-pair relevance values generated by the ProtXAI XAI analysis are not included in this deposition due to the size limitations. Researchers interested in accessing the relevance values should contact the authors.