Iterative machine learning-based chemical similarity search to identify novel chemical inhibitors

Table 1 Comparison of model performance for the four test targets using different chemical pairing scheme

	WEE1 (P30291)	MEK1 (Q02750)	EPHB4 (P54760)	TYR (P14679)	Avg. AUC
Number of chemical data
Known actives (< 100 nM)	19	24	22	22
New exp. data (active/inactive^a)	0/31	4/23	1/30	3/17
AUC PR
None	0.736	0.795	0.681	0.612	0.706
P_newP_prv (PP)	0.744	0.758	0.669	0.690	0.715
N_newP_prv (NP)	0.832	0.809	0.746	0.651	0.760
N_newN_prv (NN)	0.757	0.821	0.695	0.651	0.731
PP–NP	0.828	0.803	0.743	0.748	0.781
NP–NN	0.839	0.829	0.753	0.665	0.772
PP–NN	0.762	0.773	0.693	0.710	0.735
PP–NP–NN	0.839	0.815	0.745	0.742	0.785

Experimental chemical activity data and cross-validation results (AUC of Precision-Recall curve) are shown for each target. The chemical compounds are labeled as P_new (new active), P_prv (previous active), N_new (new inactive), and N_prv (previous random inactive data)
^aChemical compounds with lower than POC 20% and higher than 80% are defined as active and inactive compounds, respectively

ISSN: 1758-2946