Anisotropic Message Passing: Graph Neural Networks with Directional and Long-Range Interactions

Authors: Moritz Thürlemann, Sereina Riniker

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

Reproducibility Variable Result LLM Response
Research Type Experimental Results are split into three parts: First, two model systems are investigated to explore the proposed modifications. Second, the model is applied to an existing dataset of QM/MM systems. Finally, results for QM9 are reported as a comparison with existing models.In Table 1, results for a 3-layer AMP(3)-DP model are reported.Table 3: Mean absolute error (MAE) for the 12 properties of the QM9 dataset (Ramakrishnan et al., 2014). AMP refers to the model proposed in this work.
Researcher Affiliation Academia Moritz Th urlemann, Sereina Riniker Department of Chemistry and Applied Biosciences ETH Z urich Vladimir-Prelog-Weg 2, 8093 Z urich, Switzerland moritzt@ethz.ch, sriniker@ethz.ch
Pseudocode No The paper provides mathematical equations and descriptions of the message passing steps, but no explicitly labeled 'Pseudocode' or 'Algorithm' block, nor a structured procedure formatted like code.
Open Source Code Yes Code and model parameters used to produce the results in this work are made available at github.com/rinikerlab/AMP.
Open Datasets Yes All datasets used in this work are either already publicly available or made available under the same URL [github.com/rinikerlab/AMP].
Dataset Splits Yes The datasets were taken from B oselt et al. (2021) and randomly split into training/validation/test sets of 7 000/2 000/1 000 samples, respectively.
Hardware Specification Yes All timings are reported based on a Nvidia Titan V.The authors gratefully acknowledge NVIDIA for providing a Titan V under the NVIDIA Academic Hardware Grant Program.
Software Dependencies Yes Models were implemented with Tensor Flow (2.6.2 and 2.9.1) (Abadi et al., 2015; Developers, 2021).
Experiment Setup Yes Three message passing layers were used except for the QM9 dataset, where we used four message passing layers.Model parameters were optimized using ADAM (Kingma & Ba, 2017) with default parameters (β1 = 0.9, β2 = 0.999, ϵ = 10 7) and an exponentially decaying learning rate (5 10 4, 10 5).Gradients were clipped by their global norm with a clip norm of 1 (Pascanu et al., 2012).Single layer models were used for all experiments. Models were trained for 10 000 steps using a batch size of five.