Can human experts predict solubility better than computers?

Table 5 Performance of best individual classifiers, errors are absolute (unsigned) and are measured in log S units

Compound	MLP error	Human 11 error	Difference
4-Aminobenzoic acid	0.42	0.63	− 0.21
4-Aminosalicylic acid	0.39	0.04	0.35
Antipyrine	1.90	1.48	0.42
Chloramphenicol	0.78	0.89	− 0.11
Corticosterone	0.00	0.76	− 0.76
Dapsone	0.41	0.09	0.32
Primidone	1.45	0.36	1.09
Estrone	0.78	1.32	− 0.54
Alclofenac	0.02	1.13	− 1.11
5-Fluorouracil	0.07	0.97	− 0.90
Griseofulvin	0.90	1.25	− 0.35
Fluometuron	0.33	0.46	− 0.13
Fluconazole	0.22	0.20	0.02
Khellin	0.13	0.02	0.11
Clozapine	0.33	0.76	− 0.43
Norethisterone	1.53	0.37	1.16
Nicotinic acid	0.59	0.15	0.44
Perphenazine	0.53	0.84	− 0.31
Pteridine	1.00	0.02	0.98
Salicylamide	0.23	1.34	− 1.11
Sulfanilamide	0.71	0.14	0.57
Gliclazide	1.30	0.29	1.01
Trihexyphenidyl	2.93	2.20	0.73
Triphenylene	0.38	0.73	− 0.35
Mifepristone	0.86	1.90	− 1.04
Average	0.728	0.734	− 0.005

The difference is meaningfully signed, with a positive value where the best human classifier performed better on that compound and a negative value where the best machine learning classifier performed better

ISSN: 1758-2946