Table 1 Comparison of CycleGPT-HyperTemp and other models

From: Exploring the macrocyclic chemical space for heuristic drug design with deep learning models

Methods

Validity (%)

Macrocycle_ratio (%)

Novel_unique_macrocycles (%)

HMM

3.29 ± 0.030

0.18 ± 0.021

0.18 ± 0.021

N_gram

10.35 ± 0.035

2.85 ± 0.029

2.57 ± 0.013

Char_RNN

56.37 ± 0.074

56.15 ± 0.056

11.76 ± 0.022

AAE

14.82 ± 0.265

13.00 ± 0.137

10.86 ± 0.094

VAE

22.31 ± 0.183

20.19 ± 0.123

14.14 ± 0.250

ORGAN

6.46 ± 0.151

0 ± 0

0 ± 0

MolGPT

100 ± 0

0 ± 0

0 ± 0

Llamol

76.10 ± 0.209

75.29 ± 0.192

38.13 ± 0.125

MTMol-GPT

71.95 ± 0.097

70.52 ± 0.030

31.09 ± 0.199

cMol-GPT

7.87 ± 0.301

6.26 ± 0.209

6.25 ± 0.197

CycleGPT-HyperTemp

79.02 ± 0.017

75.98 ± 0.002

55.80 ± 0.002

  1. For each method, the mean ± SD of the evaluated metrics are reported for three parallel runs, each sampling 30000 molecules to compensate for random seed fluctuations. The three metrics are all percentages indicating the ratio of the corresponding data to the total number of samples. CycleGPT-HyperTemp is based on the proposed CycleGPT generative model with the HyperTemp sampling scheme in the inference stage.
Back to article page