A multitask GNN-based interpretable model for discovery of selective JAK inhibitors

Table 1 R² performance of each model on the four tasks

	Methods	Training set	Validation set	Test set
Global	MTATFP	0.96	0.79	0.78
	STATFP	0.93	0.75	0.73
	LightGBM_MD	0.93	0.69	0.70
	LightGBM_ECFP4	0.91	0.71	0.70
JAK1	MTATFP	0.95	0.80	0.82
	STATFP	0.95	0.80	0.80
	LightGBM_MD	0.91	0.75	0.72
	LightGBM_ECFP4	0.91	0.75	0.76
JAK2	MTATFP	0.97	0.83	0.81
	STATFP	0.96	0.82	0.8
	LightGBM_MD	0.94	0.70	0.75
	LightGBM_ECFP4	0.92	0.75	0.71
	Xgboost^a	0.97^a	0.80^a	0.80^a
JAK3	MTATFP	0.97	0.77	0.76
	STATFP	0.91	0.68	0.70
	LightGBM_MD	0.92	0.69	0.70
	LightGBM_ECFP4	0.90	0.69	0.73
TYK2	MTATFP	0.94	0.76	0.75
	STATFP	0.91	0.69	0.63
	LightGBM_MD	0.93	0.61	0.62
	LightGBM_ECFP4	0.91	0.65	0.61

It showed that the performance of various models (deep learning methods based on MTATFP or STATFP strategies and LightGBM-based machine learning approaches) on each task. The closer the R² value was to 1, the better the model performed
^ais the best results in Yang’s work. Although each dataset could not be guaranteed to be the identical, our multitasking model has obvious advantages as well

ISSN: 1758-2946