(in bold=engineering problem, non-bold=science problem)

Pain points

Current pain points with setup and analysis pipelines:

• Everything is scripts
• The scripts are very brittle
• Long-term: calculations should be run automatically when designs are submitted to postera (in a specific way)
  • Predictions should also be appended back onto the designs on postera

Setup pain points:

• A human needs to find the best reference to use (e.g. in fragalysis)

• A human needs to identify common core and reference structures for docking ( FEGrow)

• Docking/posing is suboptimal because related ligands often have different poses, resulting in geometry-derived atom maps that cause transformations to be larger than necessary ( FEGrow)

• we’re not using any compchem modelling to suggest new designs ( FEGrow)

• Network planning is rudimentary (star or multi-star maps) (OpenFE)

  • we don’t currently try to minimize number of charge changes
  • can fairly easily adopt a LOMAP/HIMAP approach

• Current geometry-derived atom mapping performance is suboptimal and fragile (OpenFE)

  • We don’t have sufficient redundancy built into maps to deal with poor or difficult edges (such as large changes or charge changes)
  • We don’t use any information from bound equilibrium simulations in assessing similarity between poses or which parts of the ligand scaffold are most stable and which parts are flexible or conformationally heterogeneous

• Long-term we would want to work with a ‘project-wide’ FE network. This means that whatever new ligands are submitted for FECs, the optimal location in the network is found per ligand, then the new edges are run. The statistics (edge uncertainties and any available experimental information) across the whole network are used in the FE calculation. Along the course of a LeadOpt project this would result in increasingly reliable predictions.

  • this would also be a great paper because we can publish all ligands iterated in our LeadOpt projects
  • Alternatively we could use a NetBFE-type approach but this has its own shortcomings

• We don't have an automatic way of insetting (chemical) intermediates that would greatly reduce the complexity and cost of alchemical networks

Production:

• We can’t easily tell when an edge has not converged and should be omitted (or run for a bit longer)