Table 2.

Comparison architectures A, B, C and D

Architecture	Merge mode	Layer count	Layer size	Best model epoch#	Validity%	Uniqueness%	Training%	Length match%a	HAC match%b
A: LSTM–LSTM	–	1/1	64/64	54, 72, 63	95.4 ± 0.4	99.9 ± 0.1	12.0 ± 0.9	98.2 ± 0.3	94.0 ± 0.9
B: biLSTM–biLSTM	–	1/1	64/64	20, 22, 28	96.5 ± 0.5	99.9 ± 0.1	12.5 ± 0.9	97.9 ± 0.5	94.9 ± 0.8
A: LSTM–LSTM	–	1/1	256/256	17, 17, 20	96.7 ± 0.4	99.9 ± 0.1	15.0 ± 0.7	98.2 ± 0.9	94.0 ± 1.8
B: biLSTM–biLSTM	–	1/1	256/256	6, 7, 10	97.1 ± 0.4	99.9 ± 0.1	13.1 ± 0.5	98.2 +/0.6	93.9 ± 0. 8
C: biLSTM–biLSTM	Concatenated	1/4	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	Average	1/4	64/64	11, 15, 15	97.2 ± 0.3	99.9 ± 0.1	12.5 ± 0.3	98.6 ± 0.2	96.1 ± 0.7
C: biLSTM–biLSTM	Learnable average	1/4	64/64	15, 17, 23	97.6 ± 0.2	99.9 ± 0.0	14.6 ± 0.2	97.4 ± 0.4	94.8 ± 1.2
D: biLSTM–biLSTM	Concatenated	4/4	64/64	11, 11, 9	96.9 ± 0.3	99.9 ± 0.0	14.4 ± 0.5	97.4 ± 0.2	95.6 ± 1.2
D: biLSTM–biLSTM	Average	4/4	64/64	15, 17, 14	96.7 ± 0.1	99.9 ± 0.0	11.9 ± 0.2	98.1 ± 0.5	95.3 ± 1.1
D: biLSTM–biLSTM	Learnable average	4/4	64/64	12, 25, 18	95.6 ± 0.1	99.9 ± 0.0	10.4 ± 0.5	98.0 ± 0.2	96.2 ± 0.6
Influence of bidirectionality
LSTM–LSTM	Concatenated	1/4	64/64	20,17,31	96.8 ± 0.4	99.9 ± 0.1	13.4 ± 0.5	97.6 ± 0.8	94.8 ± 1.3
biLSTM-LSTM	Concatenated	1/4	64/64	9, 14, 9	97.1 ± 0.3	99.9 ± 0.1	13.2 ± 0.5	97.7 ± 0.9	95.5 ± 1.4

Best architecture is highlighted in italics

^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”)