NEVAE: A Deep Generative Model for Molecular Graphs

Authors: Bidisha Samanta, Abir De, Gourhari Jana, Vicenç Gómez, Pratim Chattaraj, Niloy Ganguly, Manuel Gomez-Rodriguez

JMLR 2020 | Venue PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Experiments reveal that our variational autoencoder can discover plausible, diverse and novel molecules more effectively than several state of the art models.
Researcher Affiliation Academia 1Department of Computer Science and Engineering, IIT Kharagpur, Kharagpur, India 2Department of Computer Science and Engineering, IIT Bombay, Mumbai, India 3Department of Chemistry, IIT Kharagpur, Kharagpur, India 4DTIC, Universitat Pompeu Fabra, Barcelona, Spain 5Max Planck Institute for Software Systems, Kaiserslautern, Germany
Pseudocode Yes Algorithm 1: PROPERTYORIENTEDDECODER: it trains a parameterized propertyoriented decoder.
Open Source Code Yes To facilitate research in this area, we are releasing an open source implementation of our model in Tensorflow as well as synthetic and real-world data used in our experiments1. 1. https://github.com/Networks-Learning/nevae
Open Datasets Yes We experiment with molecules from two publicly available datasets, ZINC (Irwin et al., 2012) and QM9 (Ramakrishnan et al., 2014).
Dataset Splits Yes More specifically, we first sample 3,000 molecules from our ZINC dataset, which we split into training (90%) and test (10%) sets.
Hardware Specification Yes We carried out all our experiments for NEVAE using Tensorflow 1.4.1, on a 64 bit Debian distribution with 16 core Intel Xenon CPU (E5-2667 v4 @3.20 GHz) and 512GB RAM.
Software Dependencies Yes We carried out all our experiments for NEVAE using Tensorflow 1.4.1, on a 64 bit Debian distribution with 16 core Intel Xenon CPU (E5-2667 v4 @3.20 GHz) and 512GB RAM.
Experiment Setup Yes Therein, we had to specify four hyperparameters: (i) D the dimension of zu, (ii) K the maximum number of hops used in encoder to aggregate information, (iii) L the number of negative samples, (iv) lr the learning rate. Note that, all the parameters W s and b s in the input, hidden and output layers depend on D and K. We selected these hyperparameters using cross validation. More specifically, we varied lr in a logarithmic scale, i.e., {0.0005, 0.005, 0.05, 0.05}, and the rest of the hyperparameters in an arithmetic scale, and chose the hyperparameters maximizing the value of the objective function in the validation set. For synthetic (real) data, the resulting hyperparameter values were D = 7(5), K = 3(5), L = 10(10) and lr = 0.005(0.005). To run the baseline algorithms, we followed the instructions in the corresponding repository (or paper). ... We implemented stochastic gradient descent (SGD) using the Adam optimizer.