examples

Structure in config.toml

The config file is structured in different sections. The parameters relevant for this work are described. Information about other parameters and options can be found in the repository of REINVENT4.

General settings

• run_type: Set to staged_learning to run the inverse design for molecular generation guided by reward function.
• device: Set to cpu or cuda:0 depending on the available hardware.

Parameters section

• summary_csv_prefix : Prefix for the summary csv file containing the results of the run.
• prior_file: Path to a pretrained prior model. The priors used in this work can be found in the REINVENT4/priors/ directory. For this work the reinvent.prior, reinvent_zinc.prior, and reinvent_tadf.prior models were used.
• agent_file: Path to the pretrained agent model. Generally, prior and agent refer to the same model, however, prior is a model that is trained to generate molecules according to some training distribution and agent is a fine-tuned model that is trained to generate molecules according to some property distribution defined by the reward function.
• batch_size: Number of molecules generated per batch.

Diversity filter section

This set of parameters is optional and can be used to increase the diversity of generated molecules.

• type: Choose between IdenticalTopologicalScaffold, ScaffoldSimilarity, PenalizeSameSmiles, and IdenticalMurckoScaffold. In this work, ScaffoldSimilarity was used exclusively.
• bucket_size: Number of molecules with the same scaffold/similarity that can be accepted before penalizing further molecules.
• minscore: Minimum score a molecule must have to be registered in the diversity filter.
• minsimilarity: Minimum similarity threshold for scaffold similarity.
• penalty_multiplier: penalty factor for PenalizeSameSmiles

Scoring settings

• chkpt_file: Name of the checkpoint file to store the agent model after the run.

• max_score: Termination criterion for the run.

• min_steps: Minimum number of steps to run.

• max_steps: Maximum number of steps to run.

• type: Choose the aggregation method for the individual components of the reward function. Available options are custom_product, custom_sum, hypervolume, and prod_plus_hypervolume. For the aggregation methods using the hypervolume its important that the individual scoring components are scaled between 0 and 1 and that all components have the same weight.

Scoring components

• Custom alerts
  The custom alerts component can be used to penalize molecules containing unwanted substructures.

  • name: Set to Unwanted SMARTS
  • weight: weight to fine-tune the relevance of this component
  • params.smarts: List of SMARTS strings defining unwanted substructures.

• Triplet energy prediction
  The EnTdecker model is used to predict the triplet energy of generated molecules as described in this paper.

  • name: Set to name for scoring component, e.g., "EnTdecker"
  • weight: set weight of the component in the overall score.
  • params.checkpoint_dir: Path to the directory containing the pretrained EnTdecker model. A downloaded model can be found in Models/triplet_energy/model_42.pt
  • params.rdkit_2d_normalized: Set to true to use normalized 2D descriptors. Required for the EnTdecker model.
  • params.target_column: Set to "e_t". Required for the EnTdecker model.

• ML-predicted absorption wavelength prediction
  The multi-fidelity model described by Greenman et al. paper is used to predict the maximum absorption wavelength of generated molecules.

  • name: Set to name for scoring component, e.g., "ChemProp_uvvis"
  • weight: set weight of the component in the overall score.
  • params.checkpoint4featuregen: Path to ensemble of ChemProp models used for predicting S1 excitation energy. The models used in this work can be found in Models/uvvis/lambda_max_abs_wb97xd3/chemprop/all_wb97xd3/production/fold_0
  • params.checkpoint_dir: Path to the ChemProp model checkpoint used for the low-fidelity prediction. The model used in this work can be found in Models/uvvis/lambda_max_abs/chemprop_tddft/combined/production/fold_0
  • params.tmp_dir: Path to a temporary directory for storing intermediate files.
  • params.target_column: Set to "peakwavs_max".
  • params.rdkit_2d_normalized: Set to false.

• Semi-empirical absorption wavelength prediction
  The semi-empirical excited state calculation component uses xtb and stda to calculate the maximum absorption wavelength of generated molecules.

  • name: Set to name for scoring component, e.g., "SQM_lambda_max"
  • weight: set weight of the component in the overall score.
  • params.tmp_dir: Name of a temporary directory for storing intermediate files. Since this folder gets created temporarily for intermediate files and deleted afterwards, it must be a non-existing directory. So, e.g., /InvEnT/tmp/multiwfn_calcdir where the directory multiwfn_calcdir gets created during the generative process and can not be present before running the code. The run will terminate if the provided path is an existing directory to prevent unintential deletion of folder.
  • params.path_to_xtb: Path to the xtb executable.
  • params.path_to_stda: Path to the directory containing the xtb4stda binary.
  • params.maximum_waiting_time: Maximum waiting time for the geometry optimization in seconds.
  • params.use_stddft: Set to false to use stda instead of stddft for excited state calculations.
  • params.use_gfnff: Set to true to use gfn-ff for geometry optimization instead of gfn2.
  • params.target_property: Set to lambda_max.

• Excited state character
  Computes the HOMO-LUMO overlap and estimates the nature of the excited state (CT or LE) using xtb and Multiwfn.

  • name: Set to "Overlap_quick".

  • weight: Set weight of the component.

  • params.dir4tempfiles:Name of a temporary directory for storing intermediate files. Since this folder gets created temporarily for intermediate files and deleted afterwards, it must be a non-existing directory. So, e.g., /InvEnT/tmp/multiwfn_calcdir where the directory multiwfn_calcdir gets created during the generative process and can not be present before running the code. The run will terminate if the provided path is an existing directory to prevent unintential deletion of folder.

  • params.path_to_xtb: Path to the xtb executable.

  • params.path_to_multiwfn: Path to the Multiwfn executable.

  • params.calculation_mode: Set to multiwfn_quick (only singlet geometry used for FMOs) or multiwfn (both singlet and triplet states optimized).

  • params.use_gfn2: Set to false (optimization uses GFN-FF).

  • params.aggregation_mode: Choose between formula (weighted sum) and threshold (binary score).

    • formula: The score is the sum of the Singlet part ($S_{\text{part}}$) and the Triplet part ($T_{\text{part}}$):
      $$\text{score} = S_{\text{part}} + T_{\text{part}}$$
      The parts are calculated using the following parameters:

      Variable	Parameter in config file	Description
      $w_{\text{singlet}}$, $w_{\text{triplet}}$	params.Singlet_param, params.Triplet_param	Weights for the overall Singlet and Triplet contributions.
      $S_{\text{overlap}}$, $S_{\text{distance}}$	params.S_overlap, params.S_distance	Weights for the singlet overlap value ($O_{S_1}$) and distance ($D_{S_1}$) of HOMO and LUMO center.
      $T_{\text{overlap}}$, $T_{\text{distance}}$	params.T_overlap, params.T_distance	Weights for the triplet overlap value ($O_{T_1}$) and distance ($D_{T_1}$) of HOMO and LUMO center.

    • threshold: The score is a weighted sum of two binary components ($S_{\text{part}}$ and $T_{\text{part}}$), which are either 0 or 1:
      $$\text{score} = w_{\text{singlet}} \cdot S_{\text{part}} + w_{\text{triplet}} \cdot T_{\text{part}}$$
      The binary parts are determined by the following conditions:

      Component	Condition	Character Represented
      $S_{\text{part}}$	1 if $O_{S_1}$ < params.S_overlap 0 otherwise.	CT (Charge Transfer)
      $T_{\text{part}}$	1 if $O_{T_1}$ > params.T_overlap 0 otherwise.	LE (Locally excited)

• Conjugation
  Computes the degree of conjugation in molecules.

  • name: Set to "Conjugation".
  • weight: Set weight of the component.
  • params.mode: Choose between fraction (fraction of conjugated bonds) and largest_conjugated_fragment (size of the largest fragment).
  • params.exclude_split_system: Set to true to exclude molecules with disconnected conjugated structures.

To define the target values for each scoring component, the corresponding transformation functions have to be set by using the transforms keywords in the config file.