This repository contains code for Pasithea, a gradient-based method for property optimization on molecules (or, if you will, a way to dream up molecules). See the corresponding scientific paper: Deep Molecular Dreaming: Inverse machine learning for de-novo molecular design and interpretability with surjective representations
At the core of model is a simple fully-connected neural network. The method outputs continuous, valid molecular transformations, which is enabled by these additions:
- Self-referencing embedded strings (SELFIES), a surjective representation for molecular graphs.
- Random noise in molecular inputs. Every zero in the one-hot encoding is replaced with a random decimal in the range [0, k], where k is a specified upperbound.
At a high-level, there are two main steps:
- Train the neural network on one-hot encoded SELFIES inputs and continuous logP outputs.
- Inverse-train the neural network. Given a single logP and a single molecule, the network generates molecular variants that are optimized toward the logP value. The weights and biases remain fixed.
The provided script demo.py may be used to view some molecular transformations and logP optimization from a subset of the QM9 dataset. settings.yml contains hyperparameters and other settings you may tweak, such as the learning rate and target logP. Note that the model is saved each time and if a model has already been trained with those hyperparameters, demo.py will load the existing model.
You may also test individual molecules by specifying them in settings.yml and running test_mols.py.
All questions and comments are welcome.
[1] Cynthia Shen, Mario Krenn, Sagi Eppel, and Alan Aspuru-Guzik. Deep Molecular Dreaming: Inverse machine learning for de-novo molecular design and interpretability with surjective representations, Machine Learning: Science and Technology 2, 03LT02 (2021).
[2] Mario Krenn, Florian Häse, AkshatKumar Nigam, Pascal Friederich, and Alan Aspuru-Guzik. Self-Referencing Embedded Strings (SELFIES): A 100% robust molecular string representation, Machine Learning: Science and Technology 1(5), 045024 (2020).