Nature Communications

Table 3 VALIDITY and CORRELATION (see the section “Evaluation metrics” for definitions) for the RNN (recurrent neural network) and Transformer models, considering similarity k₁ and k₂ trained with and without ranking loss

From: Exhaustive local chemical space exploration using a transformer model

Model	Similarity	Ranking loss	Validity	Correlation
RNN	k₁	×	0.90	0.10 ± 0.25
RNN	k₁	✓	0.71	0.22 ± 0.30
RNN	k₂	×	0.88	0.18 ± 0.24
RNN	k₂	✓	0.75	0.33 ± 0.26
Transformer	k₁	×	0.99	0.24 ± 0.17
Transformer	k₁	✓	0.99	0.63 ± 0.16
Transformer	k₂	×	0.99	0.09 ± 0.14
Transformer	k₂	✓	0.99	0.42 ± 0.22

k₁ is the Tanimoto similarity kernel, while k₂ is the kernel induced by the autoencoder. By training the models with ranking loss the correlation consistently improves, while the validity remains the same for the Transformer and decreased for the RNN. The best results are highlighted in bold.

Back to article page

Search

Advanced search

Quick links