Skip to main content

Table 7 Performance comparison of Transformer models trained on different types of molecular pairs on the Merged dataset where big property changes are desired (numbers in bracket represent the absolute increase/decrease compared to the corresponding Transformer model performance on the original test set in Table 4)

From: Transformer-based molecular optimization beyond matched molecular pairs

Test set Type of molecular pairs where Transformer is trained Successful property constraints (%) Successful structure constraints (%) Successful property and structure constraints (%)
  MMP 40.82 (\(\downarrow\) 21.08) 83.89 (\(\downarrow\) 7.66) 36.12 (\(\downarrow\) 21.97)
  Similarity (\(\ge\)0.5) 39.81 (\(\downarrow\) 12.02) 75.00 (\(\downarrow\) 7.30) 30.70 (\(\downarrow\) 13.83)
Merged Similarity ([0.5,0.7)) 38.33 (\(\downarrow\) 8.42) 66.64 (\(\downarrow\) 1.45) 25.94 (\(\downarrow\) 7.02)
  Similarity (\(\ge\)0.7) 36.14 (\(\downarrow\) 28.95) 68.57 (\(\downarrow\) 14.11) 25.58 (\(\downarrow\) 30.49)
  Scaffold 36.50 (\(\downarrow\) 25.03) 89.17 (\(\downarrow\) 6.15) 33.60 (\(\downarrow\) 23.09)
  Scaffold generic 37.78 (\(\downarrow\) 17.27) 91.30 (\(\downarrow\) 4.71) 35.26 (\(\downarrow\) 18.40)
  1. The extremes (best/worst performance or largest/smallest change) are highlighted in bold