Skip to main content

Table 3 Comparison of the percentage of generated molecules from HierG2G, Seq2Seq and Transformer that are MMPs with starting molecules

From: Molecular optimization by capturing chemist’s intuition using deep neural networks

  HierG2G Seq2Seq Transformer
  MMP_0.33 (%) MMP_0.50 (%) in Train (%) MMP_0.33 (%) MMP_0.50 (%) in Train (%) MMP_0.33 (%) MMP_0.50 (%) in Train (%)
Test-Original 52.50 61.19 45.46 73.55 83.32 43.59 90.45 96.49 48.69
Test-Molecule 25.87 30.98 54.99 67.87 78.48 51.90 88.83 95.58 56.64
Test-Property 13.04 15.20 41.49 72.32 81.77 36.35 90.69 96.37 42.02
  1. Among all the generated molecules for each test set, MMP_0.33 and MMP_0.50 represent the percentage of generated molecules that are MMPs with their corresponding starting molecules and the ratio of heavy atoms in R group to the generated molecule \(R_{group}\leq0.33\) and \(R_{group}\leq0.50\) respectively. Among all the transformations results from MMP_0.33, in Train represents the percentage of transformations that are seen in the training set
  2. For MMP_0.33 and MMP_0.50, higher values are better, and the best values are in italics