Property-Aware Relation Networks for Few-Shot Molecular Property Prediction

Authors: Yaqing Wang, Abulikemu Abuduweili, Quanming Yao, Dejing Dou

NeurIPS 2021 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Extensive experiments on benchmark molecular property prediction datasets show that PAR consistently outperforms existing methods and can obtain property-aware molecular embeddings and model molecular relation graph properly.
Researcher Affiliation Collaboration Yaqing Wang1 Abulikemu Abuduweili1,2 Quanming Yao3 , Dejing Dou1 1Baidu Research, Baidu Inc., China 2The Robotics Institute, Carnegie Mellon University, USA 3Department of EE, Tsinghua University, China
Pseudocode Yes The complete algorithm of PAR is shown in Algorithm 1.
Open Source Code Yes Codes are available at https://github.com/tata1661/PAR-NeurIPS21.
Open Datasets Yes We perform experiments on widely used benchmark few-shot molecular property prediction datasets (Table 1) included in Molecule Net [43].
Dataset Splits Yes This T is then formulated as a 2-way K-shot classification task with a support set S = {(x ,i, y ,i)}2K i=1 containing the 2K labeled samples and a query set Q = {(x ,j, y ,j)}N q j=1 containing N q unlabeled samples to be classified.
Hardware Specification No The paper states: 'Parts of experiments were carried out on Baidu Data Federation Platform.' This indicates a computing environment but does not specify any particular hardware details such as GPU models, CPU types, or memory amounts.
Software Dependencies No The paper mentions using 'RDKit [46]' for molecular graphs but does not provide a specific version number for it or any other software dependencies.
Experiment Setup Yes Appendix B: 'We train our models with Adam [48] optimizer with initial learning rate 0.001. We set the inner loop learning rate to 0.01 and outer loop learning rate to 0.001. We train the model for 100 epochs, and the batch size is 4 tasks. The number of iterations T in the adaptive relation graph learning module is set to 2.'