Dataset’s chemical diversity limits the generalizability of machine learning predictions

Table 3 Selection of functional groups detected by Checkmol with their corresponding number of molecules in PC9 and QM9 datasets and MAE in generalization conditions

Selection	Functional group classes	PC9 data		QM9 data
Selection	Functional group classes	Occurrences	QM9 model	Occurrences	PC9 model
CF and NN bonds	Azide	358	11.5	0	–
	Azo compound	272	5.3	10	4.1
	Acyl fluoride	105	6.4	0	–
	Aryl fluoride	936	9.4	1562	4.1
	Alkyl fluoride	4576	6.8	52	4.3
Abundant in QM9	Carbonitrile	4624	5.4	10315	4.7
	Secondary alcohol	6282	6.2	10668	4.3
	Trialkylamine	3301	6.4	10687	4.2
	Alkyne	3906	6.1	10873	4.5
	Tertiary amine	3388	6.4	11057	4.2
	Aromatic compound	12728	8.2	15863	4.3
	Dialkyl ether	9275	6.2	24012	4.3
	Heterocyclic compound	42665	7.0	61904	4.3
QM9 model focus	Hydroperoxide	717	45.4	0	–
	Diaryl ether	22	39.1	0	–
	Peroxide	430	32.7	0	–
	Diarylamine	11	32.6	0	–
	Carbamic acid halide	16	24.4	0	–
	Nitrite	1	22.7	0	–
	Hydroxamic acid	46	19.4	0	–
	Nitroso compound	48	17.0	0	–
	Hydroxylamine	805	16.2	13	3.4
	Hemiacetal	819	2.8	0	–
PC9 model focus	Enol ether	1977	4.4	2	16.0
	Amide acetal	59	14.2	28	7.1
	Carboxylci acid	4213	6.6	106	6.2
	Hemiaminal	3034	6.5	207	6.0
	Acyl cyanide	67	4.0	281	6.0

The selection italic CF and NN interactions, the most prominent groups in QM9, the largest MAE of the model trained on QM9 and the largest MAE of the model trained on PC9

ISSN: 1758-2946