Table 3.

Optimal number of parallel encoding layers in architecture C

Architecture	Merge mode	# Layers	Layer sizes	Best model epoch#	Validity%	Uniqueness%	Training%	Length match%a	HAC Match%b
B: biLSTM–biLSTM	–	1/1	64/64	20, 22, 28	97.1 ± 0.4	99.9 ± 0.1	13.1 ± 0.5	98.2 ± 0. 6	93.9 ± 0. 8
C: biLSTM–biLSTM	Concatenated	1/2	64/64	19, 19, 19	97.8 ± 0.4	99.9 ± 0.1	12.5 ± 0.4	97.3 ± 0.4	96.1 ± 0.1
C: biLSTM–biLSTM	Concatenated	1/3	64/64	12, 12, 12	97.2 ± 0.2	99.9 ± 0.0	12.2 ± 0.4	98.6 ± 0.3	96.9 ± 0.8
C: biLSTM–biLSTM	Concatenated	1/4	64/64	10, 14, 16	97.0 ± 0.3	99.9 ± 0.0	11.9 ± 0.6	98.5 ± 0.3	97.4 ± 0.5
C: biLSTM–biLSTM	Concatenated	1/5	64/64	8	95.9 ± 0.3	99.9 ± 0.0	13.5 ± 1.0	97.6 ± 0.2	97.2 ± 0.3
C: biLSTM–biLSTM	Concatenated	1/6	64/64	8	95.9 ± 0.2	99.9 ± 0.1	10.1 ± 0.4	96.3 ± 0.3	93.9 ± 0.7
C: biLSTM–biLSTM	Concatenated	1/7	64/64	7	96.8 ± 0.4	99.9 ± 0.0	14.0 ± 1.0	97.6 ± 0.6	95.9 ± 0.5
C: biLSTM–biLSTM	Concatenated	1/8	64/64	6, 6, 6	96.2 ± 0.7	99.9 ± 0.0	13.6 ± 0.1	98.0 ± 0.7	94.8 ± 0.8
C: biLSTM–biLSTM	Concatenated	1/16	64/64	5, 5, 5	95.9 ± 0.3	99.9 ± 0.0	13.5 ± 1.0	96.6 ± 0.7	93.1 ± 0.7

^aLength match for SMILES length distributions of the training set and generated set (See “Methods”)

^bHAC match for the atom count distributions of the generated set and training set (See “Methods”)