Table 5. Performance in validation on isolates not present in the original dataset.

RMSE (Log units)	R0 2	Correctly Classified Percentage	Overpredicted Percentage	Underpredicted Percentage	Grouping
0.54 (±0.28)	0.56 (±0.27)	0.85 (±0.13)	0.09 (±0.11)	0.06 (±0.09)	RefID (average)
0.45 (±0.33)	0.62 (±0.34)	0.84 (±0.24)	0.11 (±0.20)	0.06 (±0.15)	IsolateName (average)
0.44 (±0.34)	0.62 (±0.34)	0.84 (±0.24)	0.10 (±0.20)	0.06 (±0.15)	SeqID (average)
0.54 (±0.18)	0.58 (±0.19)	0.83 (±0.10)	0.11 (±0.10)	0.06 (±0.06)	Drug (average)
0.44	0.75	0.85	0.11	0.03	PI (Class)
0.43	0.74	0.86	0.10	0.04	ATV
0.37	0.75	0.72	0.28	0.00	IDV
0.39	0.83	0.91	0.03	0.06	LPV
0.44	0.76	0.9	0.05	0.04	NFV
0.44	0.78	0.91	0.05	0.03	RTV
0.49	0.75	0.88	0.07	0.05	SQV
0.52	0.38	0.70	0.20	0.02	TPV
0.68	0.65	0.89	0.05	0.06	NNRTI (Class)
0.64	0.63	0.83	0.10	0.07	DLV
0.60	0.70	1.00	0.00	0.00	EFV
0.76	0.65	0.87	0.04	0.09	NVP
0.61	0.39	0.79	0.12	0.09	NRTI (Class)
0.47	0.49	0.85	0.09	0.07	ABC
0.90	0.56	0.84	0.09	0.07	AZT
0.41	0.37	0.64	0.12	0.23	D4T
0.42	0.35	1.00	0.00	0.00	DDC
0.39	0.41	0.74	0.17	0.10	DDI
1.01	0.66	0.85	0.00	0.15	FTC
0.44	0.12	0.66	0.30	0.04	TDF
0.53	0.65	0.84	0.10	0.06	Overall

Validation parameters were calculated using different forms of grouping to give an unbiased error estimate. Class wide values are indicated in italic and the global average performance is indicated in bold and italic. For larger groups (RefID, SeqID, Isolatename and per drug) the average value and standard deviation are given. For three drugs (RTV, DLV, DDC) no Virco cut-off was available, here the Stanford cut off was used for both, for SQV no Stanford cut-off was available so the Virco cut-off was used for both. The table shows that our PCM models perform robustly in predicting the Log FC as indicated by the regression validation parameters RMSE and R₀ ². More importantly, the correctly classified percentage is 84% overall.