RGFN: Synthesizable Molecular Generation Using GFlowNets

Authors: Michał Koziarski, Andrei Rekesh, Dmytro Shevchuk, Almer van der Sloot, Piotr Gaiński, Yoshua Bengio, Chenghao Liu, Mike Tyers, Robert Batey

NeurIPS 2024 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We experimentally evaluate RGFN on a set of diverse screening tasks, including pretrained proxy models and GPU-accelerated docking.
Researcher Affiliation Academia 1 Mila Québec AI Institute, 2 Université de Montréal, 3 University of Toronto, 4 Jagiellonian University, 5 Mc Gill University, 6 The Hospital for Sick Children Research Institute, 7 Acceleration Consortium
Pseudocode No The paper describes the generation process in a step-by-step manner (Section 3.2) and illustrates it in Figure 1, but does not present it as formal pseudocode or an algorithm block.
Open Source Code Yes Source code available at https://github.com/koziarskilab/RGFN.
Open Datasets Yes This includes proxy models... first, the commonly used s EH proxy as described in [4]. Second, a graph neural network trained on the biological activity classification task of senolytic [74] recognition. Third, the Dopamine Receptor D2 (DRD2) oracle [53] from Therapeutics Data Commons [28]. Pretraining was done in an unsupervised fashion on the ZINC15 dataset [66].
Dataset Splits No The paper describes the training of the RGFN model using trajectory balance loss, but does not specify traditional train/validation/test dataset splits as it operates in a reinforcement learning setup against a reward function (oracle models).
Hardware Specification Yes Evaluating fragment scaling took approximately 800 GPU hours on Ge Force RTX 4090 in total. Remaining experiments took roughly 24 GPU hours per run on Quadro RTX 8000 for s EH, DRD2 and senolytic proxies, and roughly 72 GPU hours per run on an A100 for docking-based proxies.
Software Dependencies Yes We use the GPU-accelerated Vina-GPU 2.1 [68] implementation of the Quick Vina 2 [2] docking algorithm to calculate docking scores directly in the training loop of RGFN.
Experiment Setup Yes Both RGFN and FGFN were trained with trajectory balance loss [45] using Adam optimizer with a learning rate of 1 10 3, log Z learning rate of 1 10 1, and batch size of 100. The training lasted 4,000 steps. A random action probability of 0.05 was used, and RGFN used a replay buffer of 20 samples per batch. Both methods use a graph transformer policy with 5 layers, 4 heads, and 64 hidden dimensions. Exponentiated reward R(x) = exp(β score(x)) was used, with β dependent on the task: 8 for s EH proxy, 0.5 for senolytic proxy, 48 for DRD2 proxy, and 4 for all docking runs.