andyj10224 / reinvent

The application in this repository is described in the paper: "REINVENT 2.0 – an AI tool for de novo drug design"

Usage

1. Templates for inputs are provided in reinvent/data/examples/templates folder. Concrete examples for a quick start are provided in ReinventCommunity.
2. There are templates for 6 running modes. Each running mode can be executed by "python input.py some_running_mode.json" after activating the environment. Templates have to be edited before using. The only thing that needs modification for a standard run are the file and folder paths. Most running modes produce logs that can be monitored by tensorboard, see below.
   • Logging folder is defined by setting a valid path to the "logging_path" field in json. This is required for all running modes.
3. Running modes:
   • Sampling: sampling.json can be used to start sampling. Requires a generative model as an input and produces a file that contains SMILES. We provide a generative model "reinvent/data/augmented.prior". Alternatively focused Agents generated by transfer learning or reinforcement learning can be sampled as well. Example
   • Transfer Learning (TL): transfer_learning.json is the relevant template and it can be used to focus the general prior towards a narrow chemical space by training on a representative sample of SMILES provided by user. Requires as an input a list of SMILES (example format in "reinvent/data/smiles.smi") and the generative model "reinvent/data/augmented.prior". The result will be a set of generative Agent checkpoints produced after each epoch of training and a final focused Agent. Inspect the tensorboard logs to estimate which Agent has the level of focusing that you prefer. Example
   • Reinforcement Learning (RL): Use reinforcement_learning.json as a template. The general input requires paths for both Agent and Prior generative models (in "reinforcement_learning" section of the JSON files). Both can be the same model provided by us "reinvent/data/augmented.prior" or alternatively the user can provide a focused Agent generated by TL. The output is a focused generative model and "scaffold_memory.csv" file which contains the best scoring SMILES during the RL run. The output folder is defined by setting a value for "resultdir". The scoring function object "scoring_function" can be either "name": "custom_product" or "name": "custom_sum". Scoring function has a list of parameters "parameters":[] which may contain any number of component objects. The current template example offers 5 components: a QED score, Matching Substructure (MS), Custom Alerts (CA) and 2 Predictive Property (PP) components. The PP components require setting either a classification ("reinvent/data/drd2.pkl") or regression ("reinvent/data/Aurora_model.pkl") model paths. Example 1, Example 2 , Example 3.

Tutorials / jupyter notebooks

There is another repository containing useful jupyter notebooks related to REINVENT called ReinventCommunity. Currently, it contains notebooks to illustrate the Reinforcement Learning mode and the score transformation functionality (including how to find appropriate parameters for a component). Note, that the latter uses a different conda environment to execute.

Available components

The scoring function is built-up from components, which together define the "compass" the Agents use to navigate the chemical space and suggest chemical compounds. Currently, there are the following components available:

• PREDICTIVE PROPERTY: Descriptor-based models to predict e.g. activity against a given target or solubility. Uses scikit-learn models (interface). Works with both classification and regression models.
• TANIMOTO SIMILARITY: Requires a user defined set of SMILES and returns the highest similarity score to the provided set.
• JACCARD DISTANCE: Requires a user defined set of SMILES and returns the lowest distance score to the provided set.
• MATCHING SUBSTRUCTRE: This is a penalty component to bias towards generating certain (sub-)structures. Requires a user defined set of SMARTS. Returns 1 if there is a substructure match and 0.5 otherwise.
• CUSTOM ALERTS: This is a penalty component to avoid generating certain (sub-)structures. Requires a user defined set of SMARTS patterns indicating unwanted moieties. Returns 0 if there is a match and 1 otherwise.
• QED SCORE: Uses the QED implementation in RDKitlink.
• MOLECULAR WEIGHT: Physico-chemical property calculated by RDKitlink.
• TPSA: Physico-chemical property calculated by RDKitlink.
• ROTATABLE BONDS: Physico-chemical property calculated by RDKitlink.
• NUMBER OF HYDROGEN BOND DONORS: Physico-chemical property calculated by RDKitlink.
• NUMBER OF HYDROGEN BOND ACCEPTORS: Physico-chemical property calculated by RDKitlink.
• SLOGP: Atom-based calculation of LogP using Crippen’s approach as implemented in RDKitlink.
• NUMBER OF RINGS: Physico-chemical property calculated by RDKitlink.
• SELECTIVITY: If the aim is to optimize activity against one target while reducing activity agains another, i.e. to increase compound's selectivity, this component can be used. Uses two scikit-learn models. Works with both classification and regression models. One model is predicting the target activity and the other is providing an off-target prediction. The score is reflecting a user defined activity gap between the target and the off-target predictions.

---

To use tensorboard for logging:

1. To launch tensorboard, you need a graphical environment. Write: tensorboard --logdir "path to your log output directory" --port=8008 This will give you an address to copy to a browser and access to the graphical summaries from tensorboard.

2. Further command-line parameters can be used to change the amount of scalars, histograms, images, distributions and graphs shown, e.g.: --samples_per_plugin=scalar=700, images=20

Installation

1. Install Anaconda / Miniconda

2. Clone the repository

3. Open terminal, go to the repository and generate the appropriate environment: conda env create -f reinvent_shared.yml

4. (Optional) To set environmental variables (currently not needed), for example a license: On the command line, first:

   cd $CONDA_PREFIX
   mkdir -p ./etc/conda/activate.d
   mkdir -p ./etc/conda/deactivate.d
   touch ./etc/conda/activate.d/env_vars.sh
   touch ./etc/conda/deactivate.d/env_vars.sh
   

   then edit ./etc/conda/activate.d/env_vars.sh as follows:

   #!/bin/sh
   export SOME_LICENSE='/path/to/your/license/file'
   

   and finally, edit ./etc/conda/deactivate.d/env_vars.sh :

   #!/bin/sh
   unset SOME_LICENSE
   

5. Activate environment: conda activate reinvent_shared.v2.1

6. (Optional) In the project directory, in ./configs/ create the file config.json by copying over example.config.json and editing as required. In the current version this is only relevant for the unit tests.

7. Use the tool.

Tests - currently the number of tests is significantly reduced for this repo.

Before running unit tests make sure to set up your config.json. Use the provided configs/example.config.json, rename it to configs/config.json and update the USER_NAME and MAIN_TEST_PATH variables. This file will not be version-controlled. The tests can be executed using Unittest:

python main_test.py