Retrosynthetic-Reaction-Prediction

2019"默克"杯逆合成反应预测大赛

Competition URL

Model: Transformer/RNN

Framework: tensor2tensor/opennmt-py

Data format: SMILES

Visualize Tool: RDKit

Evaluation:

Evaluation Method is the same as SQuAD dataset, code here

Data from 《Retrosynthetic Reaction Prediction Using Neural Sequence to Sequence Models》

modelName	framework	score	f1	em
transformer-base	tensor2tensor	0.627	0.764	0.218
transformer-merge	tensor2tensor	0.636	0.770	0.235
transformer-base	opennmt-py	0.646	0.860	0.002
transformer-base-200000	opennmt-py	0.650	0.865	0.004
transformer-base-254000	opennmt-py	0.653	0.869	0.005
transformer-base-400000	opennmt-py	0.660	0.877	0.007
transformer-base-800000	opennmt-py	0.664	0.881	0.013
rnn-based-115000	opennmt-py	0.415	0.554	0.000

Data from "默克杯"比赛:

modelName	framework	metric	score	f1	em
rnn-based-6000	opennmt-py	char-based	0.500	0.663	0.000
rnn-based-48000	opennmt-py	char-based	0.505	0.673	0.0015
rnn-based-48000	opennmt-py	reactants-based	0.0098	0.0126	0.0015

Useful Paper and Code:

《Retrosynthetic Reaction Prediction Using Neural Sequence-to-Sequence Models》

《Found in Translation Predicting Outcomes of Complex Organic Chemistry Reactions using Neural Sequence to Sequence Models》