@misc{oai:repo.qst.go.jp:00084371,
 author = {Miho, Shidahara and Chie, Seki and Hiroshi, Watabe and Miho, Shidahara and Chie, Seki},
 month = {Oct},
 note = {Purpose: The purpose of this study was the proposal of merging biomathematical model and machine learning approach to predict pharmacokinetics, time-activity curves (TACs), of candidate PET radioligand in brain for the development of PET CNS radioligands.

Methods: Biomathematical model used in this study was based on the simplified one-tissue compartment model (1TCM) with the kinetic parameters (K1, k2 and BPND) in the human brain (Guo et al., JNNM, 2009). In silico apparent volume (Vx), lipohilisity (MlogP) of the ligand, in vitro affinity of the ligand (KD) to the target molecule and physiological parameter, the density of molecular target (Bmax) were used for the prediction of kinetic parameters (K1, k2 and BPND) (Arakawa et al., JNM, 2017). TAC can be calculated using K1, k2 and BPND and common arterial input function Cp (t). 

For merging this biomathematical model and machine learning, random forest (RF) algorithm, was introduced. As the training, 28 CNS radioligand database (Guo et al., JNNM, 2009), which includes mature, under developing and failed PET radioligands for various imaging targets was used. In total 280 datasets (28 CNS radioligand with 10 of Bmax values) was numerically created by biomathematical model. Input data for the prediction model was Vx, KD, Bmax, and MlogP, and the output was set to radioactive concentration [kBq/ml]. 3 PET radioligands ([11C]PIB, [11C]BF227, [18F]FACT)were used for the prediction of TACs. To verify the predicted radioactivity concentration, both predicted TACs from merged approach and from biomathematical model only were compared against averaged TAC from clinical PET study. 

Results: Correlation coefficient (R2) between training data and predicted radioactivity concentration [kBq/ml] was 0.97, 0.96, and 0.96 for time 5, 30 and 60 min, respectively. For the prediction of TACs for 3 PET radioligands, the merging approach resulted in poor prediction of TACs (different shape of TACs from clinical data) especially [18F]FACT compared with that from biomathematical model only. 
There are possible reasons for observed poor prediction, the number of training data, PET radioligand for the prediction, information of input datasets. 
Conclusion: In this study, we proposed the approach merging biomathematical model and machine learning to predict time-activity curves for PET radioligand. Further optimization of machine learning, and increase of applicable datasets would be necessary., 2021 VIRTUAL IEEE NCLEAR SCIENCE SYMPOSIUM AND MEDICAL IMAGING CONFERENCE},
 title = {Merging Biomathematical modelling and Machine learning to predict time-activity curves for PET CNS radioligand development},
 year = {2021}
}