Achieving well-informed decision-making in drug discovery: a comprehensive calibration study using neural network-based structure-activity models

Table 10 Time required for models included in the study. The left column lists the time needed for training (\(T_{train}\)), and the right column the time needed for inference (\(T_{inference}\)). The meanings of the abbreviations used in the table are provided below. Typical training times for models (\(T_{train}\)) used in this study are listed in the last column

Model	\(T_{train}\)	\(T_{inference}\)	Typical Training Time
MLP	\(n_{m} * (T_{f} + T_{b})\)	\(T_{f}\)	\(\sim\) 7:30 min
MLP-D	\(n_{m} * (T_{f} + T_{b})\)	\(n_{f} * T_{f}\)	\(\sim\) 7:30 min
MLP-E	\(n_{e} * n_{m} * (T_{f} + T_{b})\)	\(n_{e} *T_{f}\)	\(\sim\) 188 min
HBLL	\(n_{m} * (T_{f} + T_{b}) + L * n_{s} * (T_{flast} + T_{s})\)	\(n_{s} * T_{f}\)	\(\sim\) 175 min

1) Required time T for...
\(T_{f}\): one forward pass, \(T_{b}\): one backpropagation step,
\(T_{s}\): one HMC step, \(T_{flast}\): forward pass of last layer
2) Number of..
\(n_{m}\): mini-batches, \(n_{f}\): forward passes in MLP-D, \(n_{e}\): base estimators
in MLP-E, \(n_{s}\): samples in HMC, L: simulation steps in HMC

ISSN: 1758-2946