Retrosynthesis Prediction with Local Template Retrieval

Authors: Shufang Xie, Rui Yan, Junliang Guo, Yingce Xia, Lijun Wu, Tao Qin

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

Reproducibility Variable Result LLM Response
Research Type Experimental In this work, we introduce Retro KNN, a local reaction template retrieval method to further boost the performance of template-based systems with non-parametric retrieval. We conduct comprehensive experiments on two widely used benchmarks, the USPTO-50K and USPTO-MIT. Especially for the top-1 accuracy, we improved 7.1% on the USPTO-50K dataset and 12.0% on the USPTO-MIT dataset. These results demonstrate the effectiveness of our method.
Researcher Affiliation Collaboration Shufang Xie1, Rui Yan1*, Junliang Guo2, Yingce Xia3, Lijun Wu3, Tao Qin3 1Beijing Key Laboratory of Big Data Management and Analysis Methods, Gaoling School of Artificial Intelligence (GSAI), Renmin University of China 2Microsoft Reserarch Aisa 3Microsoft Reserarch AI4Science {shufangxie,ruiyan}@ruc.edu.cn, {junliangguo,yingce.xia,lijuwu,taoqin}@microsoft.com
Pseudocode Yes Algorithm 1: store construction algorithm
Open Source Code No The paper does not provide an explicit statement about making the source code available or a link to a code repository for the described methodology.
Open Datasets Yes Our experiments are based on the chemical reactions extracted from the United States Patent and Trademark Office (USPTO) literature. We use two versions of the USPTO benchmark: the USPTO-50K (Coley et al. 2017) and USPTO-MIT (Jin et al. 2017).
Dataset Splits Yes The USPTO-50K contains 50k chemical reactions, split into 40k/5k/5k reactions as training, validation, and test, respectively. Meanwhile, the USPTO-MIT consists of about 479k reactions, and the split is 409k/40k/30k.
Hardware Specification Yes The last two rows present the latency with or without retrieval during inference, which are measured on a machine with a single NVIDIA A100 GPU.
Software Dependencies No The paper mentions software like 'DGL-Life Sci' and 'faiss' but does not specify their version numbers, which is necessary for reproducibility.
Experiment Setup Yes The feature extractor f is a 6-layer MPNN (Gilmer et al. 2017) followed by a single GRA layer (Chen and Jung 2021) with 8 heads. We use the hidden dimension 320 and dropout 0.2. The backbone model is optimized by Adam optimizer with a learning rate of 0.001 for 50 epochs. We also early stop the training when there is no improvement in the validation loss for five epochs. The implementation of KNN is based on the faiss (Johnson, Douze, and J egou 2019) library with Index IVFPQ index for fast embedding searching, and the K of KNN is set to 32. For the adapter network, we use the same hidden dimension as the backbone GNN. The adapter is also trained with Adam optimizer with a learning rate of 0.001.