SELFIES Generator

SELFIES-based generative RNN for molecule de novo design with property-guided generation. This project is a fork of https://github.com/alexarnimueller/SMILES_generator.

This project uses the .sfi file format for SELFIES molecular data.

Introduction

This repository uses generative recurrent neural networks (RNN) with LSTM cells to learn molecular structures represented as SELFIES (SELF-referencIng Embedded Strings). SELFIES provide a 100% robust molecular string representation — every SELFIES string decodes to a valid molecule, eliminating the invalid-molecule problem inherent to SMILES.

After training, the model can generate novel molecules. An optional property-guided generation mode (REINFORCE) steers sampling towards compounds with desired properties (e.g. high QED, target LogP, MW, TPSA).

Installation

git clone https://github.com/morawskidotmy/selfies_generator.git
cd selfies_generator
uv sync            # install all dependencies
uv sync --extra dev   # include ruff for linting

Linting

uv run ruff check .
uv run ruff format .

Usage

Training

uv run python train.py --dataset data/chembl24_10uM_20-100.csv --name test --train 20 \
    --lr 0.005 --batch 512 --after 2 --sample 100 --augment 5 --preprocess --stereo 1 \
    --val 0.1 --seed 42

With property-guided reward logging (logs mean QED each sampling round):

uv run python train.py --dataset data/chembl24_10uM_20-100.csv --name test_qed \
    --train 20 --reward qed

Sampling

uv run python sample.py --model checkpoint/test/ --out generated/test_sampled.csv \
    --epoch 9 --num 1000 --temp 1.0 --seed 42

Fine-tuning

uv run python finetune.py --model checkpoint/test/ --dataset data/actives.csv \
    --name test-finetune --lr 0.005 --epoch 19 --train 20 --sample 100 --temp 1.0 \
    --after 1 --augment 10 --batch 16 --preprocess --stereo 1 --val 0.0 --seed 42

With property-guided REINFORCE steps after fine-tuning:

uv run python finetune.py --model checkpoint/test/ --dataset data/actives.csv \
    --name test-qed-ft --reward qed --pg_steps 100 --pg_sample 64

Analysis

uv run python analyze.py --generated generated/test_sampled.csv \
    --reference data/chembl24_10uM_20-100.csv --name test --n 3 --fingerprint ECFP4

Property-Guided Generation

Available reward functions (via --reward):

Name	Description	Range
qed	Quantitative Estimate of Drug-likeness	[0,1]
logp	Gaussian reward centred on target LogP	[0,1]
mw	Gaussian reward centred on target MW	[0,1]
tpsa	Gaussian reward centred on target TPSA	[0,1]

Custom reward functions can be added in losses.py and registered in REWARD_REGISTRY.

References

• Krenn, M., Häse, F., Nigam, A., Friederich, P. and Aspuru-Guzik, A. (2020) Self-Referencing Embedded Strings (SELFIES): A 100% robust molecular string representation. Machine Learning: Science and Technology 1, 045024.
• Gupta, A., Müller, A. T., Huisman, B. J. H., Fuchs, J. A., Schneider, P. and Schneider, G. (2018) Generative recurrent networks for de novo drug design. Mol. Inf. 37, 1700111.