|
|
| The sampled stripped trajectories and intermediate data, including the trained neural network weights, are available here: covampnet_data.tar.gz (17 GB). Below is a description of the data structure. |
covampnet_data/
├── trajectories/ # MD simulations
│ ├── ZS-ab2/ # simulated system
│ │ ├──e1s1_0/ # episode 1, simulation 1
│ │ │ └── output.filtered.xtc # compressed MD simulation
│ │ ├──e1s2_0/ # episode 1, simulation 2
│ │ │ └── output.filtered.xtc
│ │ ├── ...
│ │ └── filtered.pdb # topology file
│ ├── ZS-ab3/
│ └── ZS-ab4/
├── models/ # trained models
│ ├── ZS-ab2/
│ ├── ZS-ab3/
│ └── ZS-ab4/
├── model_outputs/ # Markov state probabilities for each frame, Koopman matrices,
│ │ # Chapman-Kolmogorov tests, implied timescales, eq. distribution
│ └── ...
├── trained_model_histories/ # train/val losses of final models
│ └── ...
├── frames_for_gradient_evaluation/ # frame ids for reproduction of gradient analysis
│ └── ...
├── training_splits/ # training splits for reproduction of results
│ └── ...
└── training_seeds/ # training seeds for reproduction of results
└── ...
|
| The code and example data are available on GitHub: https://github.com/KoubaPetr/CoVAMPnet |
| Alzheimer’s disease (AD) is characterized by the deposition of misfolded tau and amyloid-beta (Aβ). Tramiprosate (TMP) and its metabolite 3-sulfopropanoic acid (SPA) are phase 3 therapeutics believed to target Aβ oligomers. It is of paramount importance to understand how TMP/SPA modulate the conformations of Aβ. Here, we studied the Aβ42 alone and in the presence of TMP or SPA by adaptive sampling molecular dynamics. To quantify the effects of drug candidates on Aβ42, we developed a novel Comparative Markov State Analysis (CoVAMPnet). First, the ensembles of learned Markov state models were aligned across different systems based on a solution to an optimal transport problem. Second, the directional importance of inter-residue distances for assignment to the Markov states was assessed by a discriminative analysis of aggregated neural network gradients. TMP/SPA shifted Aβ42 towards more structured conformations by interacting non-specifically with charged residues and destabilizing salt bridges involved in oligomerization. SPA impacted Aβ42 the most, preserving α-helices and suppressing aggregation-prone β-strands. Experimental biophysical analyses showed mild effects of TMP/SPA on Aβ42, and activity enhancement by the endogenous metabolization of TMP into SPA. The CoVAMPnet method is broadly applicable to study the effects of drug candidates on conformational behavior of intrinsically disordered biomolecules. |
Acknowledgements |