Settings guide

presto can be run directly using its CLI

presto train --parameterisation-settings.smiles "CCC(CC)C(=O)Nc2cc(NC(=O)c1c(Cl)cccc1Cl)ccn2"

or from a YAML file

presto write-default-yaml default.yaml
# Modify the yaml to set the desired smiles
presto train-from-yaml default.yaml

How to get help

For details on available options and defaults, see the settings API reference.

Running

presto train --help

will also show available options.

Note that the key option when specifying training_sampling_settings or testing_sampling_settings is sampling_protocol, which determines the available sampling settings. See the available SamplingSettings classes for a description of all implemented sampling protocols. See the associated sampling_protocol field in each class for the string identifier which should be supplied to training_sampling_settings and testing_sampling_settings fields in WorkflowSettings.

Recommended defaults

Single-molecule fit

For single molecule fits, we recommend using the default settings without modification. The default force field, AceFF-2.0, can handle charged species. By default, the mm_md_metadynamics_torsion_minimisation sampling protocol is used, which includes well-tempered metadynamics on all rotatable bonds, as well as samples generated using short MLP (and MM) minimisations. We found these sampling helpful for improving torsion scans, as they often result in configurations with erroneously large steric clashes according to the MM force field.

Congeneric series fit

For fitting congeneric series, we recommend reducing the specificity of types so that parameters are mostly shared between the common substructures in different molecules. This is intended to reduce noise in the fits by removing noise in chemically equivalent parameters resulting from the generation of different samples with molecular dynamics.

Specifically, the max_extend_distance should be changed from -1 (fully specific). We've found 2 to be a reasonable default which minimally affects the training and test loss. To run a shared fit between two TYK2 ligands, for example, generate the default yaml with

presto write-default-yaml congeneric_fit.yaml

Then modify the smiles and max_extend_distance (within type_generation_settings) options parameterisation_settings section to read:

parameterisation_settings:
    smiles:
      - CCC(CC)C(=O)Nc2cc(NC(=O)c1c(Cl)cccc1Cl)ccn2
      - CCC(=O)Nc1cc(NC(=O)c2c(Cl)cccc2Cl)ccn1
    initial_force_field: openff_unconstrained-2.3.0.offxml
    expand_torsions: true
    linearise_harmonics: true
    msm_settings:
        ml_potential: aceff-2.0
        finite_step: 0.0005291772 nm
        tolerance: 0.005291772 kcal * mol**-1 * A**-1
        vib_scaling: 0.958
        n_conformers: 1
    type_generation_settings:
        Bonds:
            max_extend_distance: 2
            include: []
            exclude: []
        Angles:
            max_extend_distance: 2
            include: []
            exclude: []
        ProperTorsions:
            max_extend_distance: 2
            include: []
            exclude:
            - '[*:1]-[*:2]#[*:3]-[*:4]'
            - '[*:1]~[*:2]-[*:3]#[*:4]'
            - '[*:1]~[*:2]=[#6,#7,#16,#15;X2:3]=[*:4]'
        ImproperTorsions:
            max_extend_distance: 2
            include: []
            exclude: []

Run this with

presto train-from-yaml congeneric_fit.yaml