Table 2.

Number of active and inactive compounds and year threshold used for the time split. ChEMBL data were temporally split into training, update1, update2 and holdout set based on the publication year. Models for the micro nucleus test and liver toxicity endpoint were trained on public data while the inhouse data were split into update and holdout set based on the internal measurement date.

Target (ID)	Training set			Update1 set			Update2 set			Holdout set
	Thresh*	Inactive	Active	Thresh*	Inactive	Active	Thresh*	Inactive	Active	Thresh*	Inactive	Active
CHEMBL220	2014	802	840	2016	211	248	2017	217	138	2020	104	113
CHEMBL4078	2014	1031	1008	2015	259	275	2016	267	202	2020	499	270
CHEMBL5763	2015	1125	600	2016	302	75	2017	307	95	2020	137	114
CHEMBL203	2012	1660	433	2014	526	213	2016	428	291	2020	341	167
CHEMBL206	2006	437	325	2012	117	63	2016	114	97	2020	158	105
CHEMBL279	2010	1955	649	2013	523	307	2014	618	137	2020	686	299
CHEMBL230	2010	475	542	2013	218	78	2015	237	80	2020	218	172
CHEMBL340	2012	1272	496	2014	439	153	2015	341	59	2020	449	107
CHEMBL240	2012	797	1938	2014	301	413	2016	265	526	2020	238	498
CHEMBL2039	2014	710	645	2015	189	192	2017	380	212	2020	134	72
CHEMBL222	2009	231	673	2011	61	227	2015	40	206	2020	74	54
CHEMBL228	2009	242	858	2011	97	373	2014	31	235	2020	79	196
Micro nucleus test	-	1475	316	2005	70	134	–	–	–	2020	98	50
Liver toxicity	-	247	445	2011	42	48	–	–	–	2020	35	15

*Thresh: Data points published (ChEMBL) or measured (micro nucleus test, liver toxicity) until this year threshold are included in the corresponding subset.