Table 2. Classification results.

For each trial shown, the models indicated by • were classified into the classes specified by ○, where W, S, H and M refer to water, sulfate, heterogen and metal, respectively. The S, H and M classes were taken as equivalent for ‘not water’. Accuracy and F ₁ scores are as in Table 2 ▸ (smallest population as the positive condition). The details of each data set can be found in Section A1.

	Input models				Scored classes
Data set	W	S	H	M	W	S	H	M	Accuracy	F 1 score
Training	2312745 solvent models from 17274 structures (0.6–5.0 Å)
1	•	•			○	○			1.00	0.98
2	•	•	•	•	○	Not water			0.99	0.92
3	•	•	•	•	○	○	○	○	0.99	0.82
4	Randomized				○	○	○	○	0.92	0.02
Validation	89824 solvent models from 2573 structures (0.7–4.3 Å)
1	•	•			○	○			1.00	0.96
2	•	•	•	•	○	Not water			0.99	0.90
3	•	•	•	•	○	○	○	○	0.99	0.81
Testing	69817 F o − F c > 3.0σ peaks from 2493 structures (0.7–4.5 Å)
1	•	•			○	○			0.99	0.98
2	•	•	•	•	○	Not water			0.99	0.90
3	•	•	•	•	○	○	○	○	1.00	0.80