Machine learning approaches to optimize small-molecule inhibitors for RNA targeting

Table 1 Computational prediction and experimental verification of the 10 newly synthesized molecules

					Predicted binding score, model
	R₁	R₂	X	Molecule (Zinc#)	Lasso^a	CNN-SMILES^a	CNN-pictorial^a	Decision tree^b	Docking values^a	Ribosome activity (%)^c	IC-50 (μM)
1	H	H	CO	32048999	− 8.53	− 7.92	− 8.18	II	− 10.75	100	No inhibition
2	H	CH₃	CH₂	22780988	− 12.67	− 10.04	− 14.75	I	− 13.77	3.07	12.6
3	H	COCH₃	CO	44442665	− 9.93	− 7.97	− 8.18	II	− 7.49	100	No inhibition
4	H	CH₂CH₃	CH₂	19595411	− 14.05	− 13.81	− 14.96	I	− 13.73	1.19	9.1
5	H	Bn	CO	2992714	− 11.16	− 9.30	− 7.33	II	− 10.43	29.10	Weak inhibitor
6	H	Bn	CH₂	22455064	− 13.91	− 13.90	− 9.89	II	− 15.29	2.21	9.77
7	H	Boc	CH₂	914572967	− 10.55	− 9.83	− 12.96	II	− 10.26	69.87	Weak inhibitor
8	H	H	CH₂	19944337	− 13.76	− 13.97	− 9.30	I	− 14.27	51.88	Weak inhibitor
9	H	COCH₃	CH₂	23218050	− 11.33	− 9.78	− 13.38	II	− 9.90	8.38	97.27
10	Cl	Boc	CO	24030014	− 9.12	− 8.11	− 8.29	II	− 7.88	35.72	Weak inhibitor
Inhibitors recall					1.00	0.75	0.50	0.75
Precision					1.00	1.00	0.50	1.00
Accuracy					1.00	0.90	0.70	0.90

^aNegative score corresponds to stronger binding; bold—effective inhibitor, italics—weak inhibitor
^bGroup classification: binding (group I) > binding (group II), groups are based on decision tree classifier (Fig. 3)
^cActivity measured in the presence of 1 mM of the compound and M. smegmatis ribosomes in a bacterial coupled transcription/translation assay

ISSN: 1758-2946