Whilst high-throughput screening (HTS) has been the starting point for many successful drug discovery programs the cost of screening, the accessibility of a large diverse sample collection, or throughput of the primary assay may preclude HTS as a starting point and identification of a smaller selection of compounds with a higher probability of being a hit may be desired. Directed or Virtual screening is a computational technique used in drug discovery research designed to identify potential hits for evaluation in primary assays. It involves the rapid in silico assessment of large libraries of chemical structures in order to identify those structures that most likely to be active against a drug target. The key question is then how many molecules do you select from your virtual screen?

The results of a virtual screening run are effectively a rank ordering of the virtual screening deck ordered by whatever scoring function(s) that have been used. The task then becomes selection of molecules for experimental determination of activity.

I posed this question on the website and the results are shown below. Whilst this obviously a limited snapshot it is interesting that there is a wide variety of responses.

Some people also emailed me with further information. For companies with large internal physical screening collections, and the ability to cherry pick samples, it effectively costs the same to fill a high density plate (>1000 compounds) as it does to select a handful of compounds. On the other hand if the scientist has to purchase compounds then the logistics and cost become a significant obstacle. It would have been interesting to compare different virtual screening techniques, academic versus biotech versus large pharma etc. but I doubt I’d get as many answers from a multi-page questionnaire.

There is an interesting publication “Predictiveness curves in virtual screening” by Charly Empereur-mot et al DOI in which they look compare several docking methods and use the predictiveness curve as a quantification of the predictive performance of virtual screening methods on a fraction of a given molecular dataset. They use the Directory of Useful Decoys datasets (DUD) for comparison and were kind enough to provide me with the results, I’ve just used the data generated using Autodock Vina.

DUD consists of a total of 2,950 active compounds against a total of 40 targets. For each active, 36 “decoys” with similar physical properties (e.g. molecular weight, calculated LogP) but dissimilar topology

As an aside the DUD dataset was designed to evaluate docking algorithms, the decoys were intentionally designed to be structurally distinct from the actives. This was done to ensure that the decoys were truly inactive. While this makes DUD-E an excellent benchmark for docking, it makes it a poor choice for machine learning. Pat Walters highlights this on his blog https://patwalters.github.io/Please-Stop-Fishing/

Compared to the typical results of high-throughput screening where the hit rate is usually <1%, as the table below shows DUD contains an unusually high concentration of actives (2-5%), but the results of the virtual screening are certainly very informative.

Target	No. of actives	No. of compounds	Prevalence
ACE	49	1846	0.0265
ACHE	107	3999	0.0268
ADA	39	966	0.0404
ALR2	26	1021	0.0255
AMPC	21	807	0.0260
AR	79	2933	0.0269
CDK2	72	2146	0.0336
COMT	11	479	0.0230
COX-1	25	936	0.0267
COX-2	426	13715	0.0311
DHFR	410	8777	0.0467
EGFR	475	16471	0.0288
ER ago	67	2637	0.0254
ER antago	39	1487	0.0262
FGFR1	120	4670	0.0257
FXA	146	5891	0.0248
GART	40	919	0.0435
GPB	52	2192	0.0237
GR	78	3025	0.0258
HIVPR	62	2100	0.0295
HIVRT	43	1562	0.0275
HMGR	35	1515	0.0231
HSP90	37	1016	0.0364
INHA	86	3352	0.0257
MR	15	651	0.0230
NA	49	1923	0.0255
P38	454	9595	0.0473
PARP	35	1386	0.0253
PDE5	88	2066	0.0426
PNP	50	1086	0.0460
PPAR	85	3212	0.0265
PR	27	1068	0.0253
RXR	20	770	0.0260
SAHH	33	1379	0.0239
SRC	159	6478	0.0245
THR	72	2528	0.0285
TK	22	913	0.0241
TRP	49	1713	0.0286
VEGFR2	88	2994	0.0294
Minimum	11	479	0.0230
Maximum	475	16471	0.0473
Mean	97	3134	0.0294
Median	50	1923	0.0265

Table 1 shows a summary of the partial metrics at 2% and 5% of the ordered dataset for virtual screens performed using Autodock Vina, partial total gain (pTG), partial area under the curve (pAUC), Enrichment factors (EF)

Table 1	Autodock Vina – Top 2% dataset					Autodock Vina – Top 5% dataset
Target	pTG 2%	pAUC 2%	EF 2%	Actives 2%	Cpds 2%	pTG 5%	pAUC 5%	EF 5%	Actives 5%	Cpds 5%
ACE	0.020	0.048	3.05	3	37	0.019	0.075	2.84	7	93
ACHE	0.024	0.038	3.74	8	80	0.019	0.107	4.11	22	200
ADA	0.020	0.000	0.00	0	20	0.018	0.000	0.00	0	49
ALR2	0.098	0.028	3.74	2	21	0.071	0.154	6.80	9	52
AMPC	0.021	0.013	2.26	1	17	0.019	0.034	0.94	1	41
AR	0.161	0.157	11.96	19	59	0.108	0.268	7.83	31	147
CDK2	0.087	0.117	9.70	14	43	0.063	0.190	5.24	19	108
COMT	0.000	0.091	4.35	1	10	0.000	0.182	5.44	3	24
COX-1	0.154	0.113	11.82	6	19	0.102	0.250	7.17	9	47
COX-2	0.322	0.234	18.03	154	275	0.193	0.397	10.14	216	686
DHFR	0.215	0.070	5.47	45	176	0.150	0.118	3.56	73	439
EGFR	0.048	0.038	3.26	31	330	0.036	0.071	2.19	52	824
ER ago	0.314	0.192	17.08	23	53	0.186	0.383	9.84	33	132
ER antago	0.059	0.110	8.90	7	30	0.040	0.173	5.08	10	75
FGFR1	0.012	0.003	0.83	2	94	0.010	0.016	0.67	4	234
FXA	0.029	0.011	1.37	4	118	0.023	0.036	1.50	11	295
GART	0.108	0.000	0.00	0	19	0.087	0.005	1.00	2	46
GPB	0.113	0.026	2.87	3	44	0.081	0.101	4.22	11	110
GR	0.023	0.099	5.72	9	61	0.019	0.111	2.55	10	152
HIVPR	0.147	0.038	4.73	6	43	0.099	0.091	3.51	11	106
HIVRT	0.047	0.121	7.95	7	32	0.038	0.161	4.14	9	79
HMGR	0.015	0.035	2.79	2	31	0.012	0.049	1.14	2	76
HSP90	0.039	0.000	0.00	0	21	0.032	0.004	0.54	1	51
INHA	0.079	0.191	12.04	21	68	0.051	0.257	6.50	28	168
MR	0.346	0.229	18.60	6	14	0.215	0.517	14.47	11	33
NA	0.019	0.000	0.00	0	39	0.018	0.000	0.00	0	97
P38	0.031	0.012	1.54	14	192	0.026	0.049	2.29	52	480
PARP	0.114	0.071	4.24	3	28	0.080	0.091	3.39	6	70
PDE5	0.047	0.009	1.68	3	42	0.037	0.043	1.81	8	104
PNP	0.011	0.000	0.00	0	22	0.009	0.000	0.00	0	55
PPAR	0.304	0.219	16.28	28	65	0.183	0.372	10.33	44	161
PR	0.012	0.009	1.80	1	22	0.010	0.027	1.47	2	54
RXR	0.653	0.330	26.47	11	16	0.362	0.620	14.81	15	39
SAHH	0.126	0.069	8.95	6	28	0.086	0.174	4.84	8	69
SRC	0.099	0.053	5.64	18	130	0.070	0.135	4.78	38	324
THR	0.129	0.097	7.57	11	51	0.091	0.149	3.87	14	127
TK	0.019	0.000	0.00	0	19	0.015	0.000	0.00	0	46
TRP	0.037	0.037	3.00	3	35	0.029	0.069	2.03	5	86
VEGFR2	0.007	0.062	4.54	8	60	0.006	0.101	2.72	12	150
Minimum	0.000	0.000	0.00	0	10	0.000	0.000	0.00	0	24
Maximum	0.653	0.330	26.47	154	330	0.362	0.620	14.81	216	824
Mean	0.105	0.076	6.20	12	63	0.070	0.143	4.20	20	157
Median	0.048	0.048	4.24	6	39	0.038	0.101	3.51	10	97

Perhaps the first thing to note is the enrichment factor (after selecting the top 2% of the dataset) over all the targets varies from 0 to a maximum of 26 with a mean of 6. Where Enrichment factors were computed as follows:

where Hits_{x %} is the number of active compounds in the top x % of the ranked dataset, Hits_t is the total number of active compounds in the dataset, N _{x %} is the number of compounds in the x % of the dataset and N _t is the total number of compounds in the dataset. Unfortunately it is not possible to predict how much enrichment might be achieved.

Another way to look is to sort the data set by score and then plot number of ligands versus the number of active identified . For DHFR active ligands were identified among the highest scoring structures, but for GART the top 40 or so scoring ligands were inactives. The diagonal line gives an idea of the prevalence of hits with random picking.

The objective for HTS Analysis is not to identify every active compound in the screening set, but rather to identify sufficient active series to support the active chemistry effort available, similarly the aim of virtual screening is not to identify every hit but rather to identify sufficient active series to support the active chemistry effort available. If we assume the percentage of true actives in the virtual library is 0.5% then the enrichment due to virtual screening might take it up to 3%. So for if you select 100 compounds for experimental determination one might expect 3 actives, if you want multiple series, (in case a series is lost due to off-target activity), you would probably want to evaluate 1000 compounds.

It is probably not wise to simply select the first 1000 compounds since it is likely that some chemotypes may be repeated, better to aim to select diverse chemotypes.

This might seem like a lot of compounds, but a back of the envelope calculation for the cost of a virtual screen is around $10,000 [taking into account hardware costs, licenses, maintenance and support, salaries], in addition you are probably going to be committing substantial biology and chemistry resources on any hits, so why would you want to penny pinch on the purchase of compounds?