Structure-Aware E(3)-Invariant Molecular Conformer Aggregation Networks

Authors: Duy Minh Ho Nguyen, Nina Lukashina, Tai Nguyen, An Thai Le, Trungtin Nguyen, Nhat Ho, Jan Peters, Daniel Sonntag, Viktor Zaverkin, Mathias Niepert

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

Reproducibility Variable Result LLM Response
Research Type Experimental 6. Experiments
Researcher Affiliation Collaboration 1Department of Computer Science, University of Stuttgart, Germany 2Max Planck Research School for Intelligent Systems (IMPRS-IS) 3German Research Center for Artificial Intelligence (DFKI) 4Department of Computer Science, Technische Universitat Darmstadt, Germany 5School of Mathematics and Physics, University of Queensland, Australia 6Department of Statistics and Data Sciences, University of Texas at Austin, USA 7Hessian.AI 8Department of Applied Artificial Intelligence, Oldenburg University, Germany 9NEC Laboratories Europe.
Pseudocode Yes Algorithm 1 Entropic FGW Barycenter, Algorithm 2 Entropic FGW with Sinkhorn projections, Algorithm 3 Stabilized LSE Sinkhorn algorithm.
Open Source Code Yes Our implementation is available at this link.
Open Datasets Yes We use four datasets Lipo, ESOL, Free Solv, and BACE in Molecule Net benchmark (Table 1) and Molecular Property Prediction Tasks We conduct our experiments on Molecule Net (Wu et al., 2018), a comprehensive benchmark dataset for computational chemistry.
Dataset Splits Yes halved using Reduce LROn Plateau after 10 epochs without validation set improvement. Table 1. Number of samples for each split on molecular property prediction, classification tasks, and reaction prediction. Lipo ESOL Free Solv BACE Co V-2 3CL Cov-2 BDE Train 2940 789 449 1059 50 (485) 53 (3294) 8280 Valid. 420 112 64 151 15 (157) 17 (1096) 1184 Test 840 227 129 303 11 (162) 22 (1086) 2366 Total 4200 1128 642 1513 76 (804) 92 (5476) 11830
Hardware Specification Yes In particular, CONAN-FGW Single-GPU CONAN-FGW on Multi-GPUs indicates the version where one and four Tesla V100-32GB are used for training, respectively.
Software Dependencies No We use efficient implementations from the RDKit package (Landrum, 2016). and Leveraging Sinkhorn iterations in our barycenter solver (Algorithm 1), we speed up the training process across multiple GPUs using Py Torch s distributed data-parallel technique. No specific version numbers for RDKit or PyTorch are given.
Experiment Setup Yes We set the size of the latent dimensions of GAT (Veliˇckovi c et al., 2018) to 128/256. and Our Sch Net configuration incorporates three interaction blocks with feature maps of size F = 128, employing a radial function defined on Gaussians spaced at intervals of 0.1 A with a cutoff distance of 10 A. and Training the entire model employs the Adam optimizer with initial learning rates selected from 1e 3, 1e 3/2, 1e 4, halved using Reduce LROn Plateau after 10 epochs without validation set improvement. and We set empirically γ in Eq.(7) is 0.2.