Summary of Pushing the Limits Of All-atom Geometric Graph Neural Networks: Pre-training, Scaling and Zero-shot Transfer, by Zihan Pengmei et al.
Pushing the Limits of All-Atom Geometric Graph Neural Networks: Pre-Training, Scaling and Zero-Shot Transfer
by Zihan Pengmei, Zhengyuan Shen, Zichen Wang, Marcus Collins, Huzefa Rangwala
First submitted to arxiv on: 29 Oct 2024
Categories
- Main: Machine Learning (cs.LG)
- Secondary: Chemical Physics (physics.chem-ph)
GrooveSquid.com Paper Summaries
GrooveSquid.com’s goal is to make artificial intelligence research accessible by summarizing AI papers in simpler terms. Each summary below covers the same AI paper, written at different levels of difficulty. The medium difficulty and low difficulty versions are original summaries written by GrooveSquid.com, while the high difficulty version is the paper’s original abstract. Feel free to learn from the version that suits you best!
| Summary difficulty | Written by | Summary |
|---|---|---|
| High | Paper authors | High Difficulty Summary Read the original abstract here |
| Medium | GrooveSquid.com (original content) | Medium Difficulty Summary A pre-trained geometric graph neural network (Geom-GNN) is used as a transferable and effective geometric descriptor for improved generalization in various applications such as drug discovery, molecular dynamics, and protein mechanism analysis. Geom-GNNs have shown promising results in predicting interatomic potential and molecular properties. However, they are often supervised on specific downstream tasks which suffer from poor generalization and performance degradation on out-of-distribution (OOD) scenarios due to the lack of high-quality data and inaccurate labels. This work explores the possibility of using pre-trained Geom-GNNs as transferable geometric descriptors for improved generalization. The scaling behaviors of Geom-GNNs under self-supervised pre-training, supervised, and unsupervised learning setups are studied. Interestingly, Geom-GNNs do not follow a predictable power-law scaling on the pre-training task, but universally lack such behavior on supervised tasks with quantum chemical labels important for screening and design of novel molecules. |
| Low | GrooveSquid.com (original content) | Low Difficulty Summary Geom-GNNs are used to represent molecular and biological systems, which helps in drug discovery, understanding protein mechanisms, and simulating molecule dynamics. However, current practices involve training Geom-GNNs on specific tasks, which can lead to poor generalization when applied to new situations. This paper explores the idea of using pre-trained Geom-GNNs as transferable geometric descriptors for better performance. The study looks at how well different Geom-GNN architectures perform under self-supervised and supervised learning setups. It finds that some Geom-GNNs are more effective than others, but they all struggle to generalize well when applied to new tasks. |
Keywords
» Artificial intelligence » Generalization » Gnn » Graph neural network » Self supervised » Supervised » Unsupervised
