Multi-Objective Molecule Generation using Interpretable Substructures

Authors: Wengong Jin, Dr.Regina Barzilay, Tommi Jaakkola

ICML 2020 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate our model on various drug design tasks and demonstrate significant improvements over state-of-the-art baselines in terms of accuracy, diversity, and novelty of generated compounds.
Researcher Affiliation Academia Wengong Jin 1 Regina Barzilay 1 Tommi Jaakkola 1 1MIT CSAIL. Correspondence to: Wengong Jin <wengong@casil.mit.edu>.
Pseudocode Yes Algorithm 1 Training method with n property constraints.
Open Source Code Yes 1https://github.com/wengong-jin/multiobj-rationale
Open Datasets Yes We pre-train all the models on the same ChEMBL dataset, which contains 1.02 million training examples. On the four-property generation task, our model is fine-tuned for L = 50 iterations, with each rationale being expanded for K = 200 times. Following Li et al. (2018), the property prediction model is a random forest using Morgan fingerprint features (Rogers & Hahn, 2010).
Dataset Splits Yes For each property, we split its property dataset into 80%, 10% and 10% for training, validation and testing.
Hardware Specification No The paper does not provide specific details about the hardware used for running the experiments.
Software Dependencies No The paper mentions RDKit but does not provide specific version numbers for any software dependencies.
Experiment Setup Yes We set the positive threshold δi = 0.5. For each positive molecule, we run 20 iteration of MCTS with cpuct = 10. Our model is fine-tuned for L = 50 iterations, with each rationale being expanded for K = 200 times.