TODO.md

In Progress

• Tests
  • Workflow test with example data
  • Trivial examples for each function
  • Unit tests for SSI
  • Unit tests for density features
• Integrate DiffNets.
  • Lay out module structure in separate branch.
  • Copy core network from DiffNets repo.
  • Try to use existing featurization.
  • Include existing DiffNets featurization and compare.
• exploratory analysis via correlation coefficients of the features
  • First tests --> not very promising.
  • Try different metric
  • Find useful application or leave it out.
• Unified tutorial in documentation. Make one page for each subpackage
  • preprocessing
    • coordinates
    • densities
  • featurization
    • structure features
    • water features
    • atom features
  • comparison
  • dimensionality reduction
  • clusters (show how to cluster on PCs)
  • SSI

Plans

• More example tcl scripts for VMD
• Facilitate calculation of JSD etc. on principal components
• Facilitate calculation of SSI on results of joint clustering.
• Feature comparison of more than two ensembles
  • with respect to the joint ensemble (all metrics)
  • with respect to a reference ensemble (will not always work for KLD)
• Use MDAnalysis instead of biotite for water featurization
• Use MDAnalysis instead of PyEMMA to read features (to avoid mmshare dependency).
• Use scikit-learn or Deeptime instead of PyEMMA for clustering.
• Use scikit-learn or Deeptime instead of PyEMMA for dimensionality reduction.
• Put shared functionality of PCA and TICA into shared functions.
• Weighted PCA/tICA? (to account for varying simulation lengths or uncertainty)
• Implement T-distributed Stochastic Neighbor Embedding (t-SNE)
  • Read up on t-SNE for molecular trajectories
  • See if we can import or adapt existing code.
  • First tests with (regular) t-SNE
  • Test time-lagged t-SNE. How to handle time-dependence across simulations/ensembles?
  • write module
  • write unit tests
• Implement a clustering algorithem designed for structural ensembles
  • Read up about CLoNe
  • First tests
  • write module
  • write unit tests
• Make file format (png/pdf?) for matplotlib optional.
• Implement Linear Discriminant Analysis.
• Implement Non-Negative Matrix Factorization.
• Implement nucleic acid torsions and pseudo-torsions, as reviewed Keating et al. and as used in x3DNA or Barnaba (Barnaba code on GitHub)

Ideas

• Logo
• Hydrogen bonds as features
• Contacts as features
  • Think about a GetContacts reader
• Position deviations as features (similar to components of RMSD)
• Estimate thresholds for significance of feature differences
  • Calculate correlation times within trajectories
  • modify p-value of KS test using correlation time
  • modify p-value of KS test using number of simulation runs per ensemble
• Wasserstein distance to compare ensembles
• Add option to whiten features
• Featurizers for other molecule types
  • ligands
  • lipids
  • nucleic acids
• Account for Bonferroni correction in comparison.
• Implement conformational entropy calculations
  • Read papers, e.g, 1, 2, 3
  • Test implementations, e.g., Xentropy to find the best way to do it.
• Implement multi-dimensional scaling
• Try to integrate functional mode analysis.
• Try to integrate VAMPnets.
• Try to integrate network analysis.
• Try to integrate Deep learning the slow modes for rare events sampling
• Check out correlationplus
• Improve clusering using Gaussian Mixture Models

Done ✓

• Colab Tutorial
  • Put Notebook on Colab and get it to run.
  • Add visualizations.
  • Fix installation via pip.
  • Fix animations (they only show white canvas).
  • Add TICA to Colab tutorial.
• Include TICA in unit tests
• Write "getting started" for documentation
• Refactoring and fixes for release 0.2
  • Restructure modules to subpackages
  • Adapt README
  • Adapt API documentation
  • Include SSI to comparison example script
  • Numbering of principal component trajectories starts with 0, should start with 1
  • Axis labels and legend name for distance matrix plot
  • Function pca_features() does not have labels
  • Function compare_projections() does not have labels or legend
• Slack channel for all developers and testers, and to provide support for the user community.
• Implement clustering in principal component space
• Option to write and load features as CSV file.

Abandoned

• Frame classification via CNN on features
  • Prototype to classify simulation frames --> Diffnets probably more powerful.
  • Interpret weights as relevance of features
  • Write module
  • Write unit tests