Sample-efficient Multi-objective Molecular Optimization with GFlowNets

Authors: Yiheng Zhu, Jialu Wu, Chaowen Hu, Jiahuan Yan, kim hsieh, Tingjun Hou, Jian Wu

NeurIPS 2023 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We empirically illustrate that HN-GFN has adequate capacity to generalize over preferences. Moreover, experiments in various real-world MOBO settings demonstrate that our framework predominantly outperforms existing methods in terms of candidate quality and sample efficiency. 5 Experiments
Researcher Affiliation Academia 1College of Computer Science and Technology, Zhejiang University 2College of Pharmaceutical Sciences, Zhejiang University 3Polytechnic Institute, Zhejiang University 4Second Affiliated Hospital School of Medicine, Zhejiang University 5School of Public Health, Zhejiang University 6Institute of Wenzhou, Zhejiang University
Pseudocode Yes A Algorithms Algorithm 1 MOBO based on HN-GFN Algorithm 2 Training procedure for HN-GFN with the hindsight-like off-policy strategy
Open Source Code Yes The code is available at https://github.com/violet-sto/HN-GFN.
Open Datasets Yes We adopt the property prediction models released by Xie et al. [61] to evaluate the inhibition ability of generated molecules against GSK3β and JNK3. We only consider two objectives: inhibition scores against glycogen synthase kinase-3 beta (GNK3β) and c-Jun N-terminal kinase-3 (JNK3) [40, 33].
Dataset Splits No The paper mentions applying 'early stopping' which implies the use of a validation set, but it does not specify the exact percentages, sample counts, or methodology for the train/validation/test dataset splits.
Hardware Specification Yes The efficiency is compared on the same computing facilities using 1 Tesla V100 GPU.
Software Dependencies No The paper mentions implementing HN-GFN in 'Py Torch [50]' but does not provide a specific version number for PyTorch or any other software dependencies.
Experiment Setup Yes Table 4: Hyper-parameters used in the real-world MOBO experiments. (Contains details such as Surrogate model Hidden size, Learning rate, Number of iterations, Early stop patience, Dropout, Weight decay, Acquisition function (UCB) β, HN-GFN Learning rate, Reward exponent, Reward norm, Trajectories minibatch size, Offline minibatch size, hindsight γ, Uniform policy coefficient, Hidden size for GFlow Net, Hidden size for hypernetwork, Training steps, α).