Chemical perception

How does this project fit into your broader strategy?

• Identifying pathological cases in the fitting procedure - finding an automated detection protocol

• Automatically determining how many parameters/types (SMIRKS patterns) are needed in a data-driven manner 

Team

Project owner: David Mobley 

Team members: Jessica Maat (Deactivated) Trevor Gokey Hyesu Jang Lee-Ping Wang Chaya Stern (Deactivated) Michael Gilson John Chodera Josh Fass (Deactivated) 

Status

IN PROGRESS

Why are we doing this?

Problem statement:

Impact of this problem:

How do we judge success?

What are possible solutions?

What do we already know?

What do we need to answer?

• Diagnostics tools - when to assign a new parameter for bonds, angles and torsions (for example, in cases of multimodal distributions per SMIRKS pattern)? How can we easily improve chemical perception? Bayesian sampling to explore the number of parameters and associated value distributions?

• How do we identify problems in particular parameters? How do we fix them?

• How much coverage of each parameter type to get good FF? 

• How do we know when our typing is bad?

• How to use WBO to improve accuracy while reducing parameter space?

• How to leverage ChemPer to address some of these questions? 

• How do we determine how many QM calculations we might need to achieve high-quality parameter coverage of a dataset like ChEMBL?

• Can Hessians identify soft degrees of freedom that should be driven (e.g. impropers, rings)?

• How can small molecule crystal structure data best be used by OpenFF?

• How do we select the right molecules to train on?

  • Parameter overrepresentation? 

  • How do we identify problems?

  • How do we fix?

• Molecule set selection and expansion: 

  • Protonation / tautomers

  • Ions (carboxylates/salt bridges, monovalent ions) - What data to use?

  • Element expansion: Si, Br, P, B

What are we doing?

Why is this a good solution?

Visualize the solution

Scale and scope