Integrating Computational Lead Optimization Diagnostics With Analog Design and Candidate Selection

Figures & data

Table 1. Analog series.

AS ID	Target name	ChEMBL Target ID	# Subst. Sites	# EAs	# FW EAs	# FW VAs
1	Serine/threonine-protein kinase mTOR	2842	2	153	27	264
2	Acetyl-CoA carboxylase 2	4829	2	112	0	218
3	Acetyl-CoA carboxylase 2	4829	6	149	33	3812
4	Adenosine A2b receptor	255	6	129	72	392
5	GABA receptor alpha-5 subunit	5112	2	193	118	1647
6	Purinergic receptor P2Y12	2001	2	237	145	2766
7	Vanilloid receptor	4794	3	162	0	0
8	Mitogen-activated protein kinase kinase kinase 12	1908389	2	111	44	1844
9	5-lipoxygenase activating protein	4550	3	259	162	5204
10	5-lipoxygenase activating protein	4550	2	100	4	96
11	Epidermal growth factor receptor erbB1	203	2	106	0	306
12	Sodium channel protein type IX alpha subunit	4296	3	146	40	2860
13	Acetyl-CoA carboxylase 2	4829	3	100	10	145
14	Proteinase activated receptor 4	4691	5	117	8	212
15	Acetyl-CoA carboxylase 2	4829	3	128	54	319
16	p53-binding protein Mdm-2	5023	4	149	72	381
17	Acetyl-CoA carboxylase 2	4829	3	116	81	331
18	Sodium channel protein type IX alpha subunit	4296	5	151	17	1736
19	P2X purinoceptor 3	2998	6	102	84	336
20	MAP kinase ERK2	4040	2	264	0	262
21	Tyrosine-protein kinase SYK	2599	5	173	43	5755
22	Prostaglandin E synthase	5658	3	168	63	3893
23	Tyrosine-protein kinase SYK	2599	3	168	72	1887
24	5-lipoxygenase activating protein	4550	2	126	0	124

The table summarizes the composition of ASs used herein and reports the proportion of existing analogs and newly generated virtual analogs that qualify for Free-Wilson potency prediction, as discussed in the text. ‘# Subst. Sites’ reports the number of substitution sites per AS.

AS: Analog series; EA: Existing analog; FW: Free-Wilson; ID: Identification; VA: Virtual analogs.

Figure 1. Principles of Free-Wilson analysis.

Shown are four analogs from the same AS that are active against GABA receptor alpha-5 subunit (AS 5; ChEMBL target ID 5112). For each compound, its logarithmic experimental potency (pK_i) value is reported. In addition, the core structure of the AS is depicted in the center and the two substitution sites R₁ and R₂ are highlighted in yellow and blue, respectively. Corresponding substituents in analogs are colored accordingly. Individual potency contributions of directed substitutions are reported as ΔpK_i values. The figure illustrates the principles of Free-Wilson predictions of compound potency.

AS: Analog series.

Figure 2. Compound optimization monitor diagnostic scores.

The scatter plot compares COMO chemical saturation (S) and SAR progression (P) scores for ASs from ChEMBL (version 25). Each dot represents a series and is scaled in size according to the number of analogs. Different combinations of S and P scores distinguish ASs at different LO stages.

AS: Analog series; COMO: Compound optimization monitor; LO: Lead optimization; SAR: Structure–activity relationship.

Figure 3. Performance of regression models.

On the vertical axis, mean coefficients of determination (R²) for regression models are reported. The horizontal axis lists ASs using their IDs according to Table 1. For each AS, two R² values are given for RR (blue) and SVR (orange) models.

AS: Analog series; ID: Identification; RR: Ridge regression; SVR: Support vector regression.

Figure 4. Comparison of Free-Wilson and support vector regression predictions.

Scatter plots compare R² values of FW and SVR predictions for individual ASs. Each dot represents an AS that is scaled in size according to the number of FW EAs. The diagonal corresponds to perfect correlation between calculated coefficients. (A) FW versus global SVR predictions (according to Figure 3). (B) FW versus SVR predictions on FW EA subsets.

AS: Analog series; EA: Existing analog; FW: Free-Wilson; SVR: Support vector regression.

Figure 5. Experimental and predicted potency values.

(A) Box plots compare experimental and predicted potency value distributions. Each triplet represents one of 11 ASs yielding predictive FW and SVR models. The y-axis reports log. potency values and the x-axis the number of FW EAs. Numbers in parentheses are AS IDs. Distributions of experimental potency values (green), mean FW predictions (red) and mean SVR predictions (orange) are reported for FW EAs. (B) Individual predictions are shown for four exemplary FW EAs (with ChEMBL IDs) from the same AS with activity against purinergic receptor P2Y12 (AS 6; ChEMBL target ID 2001). In the table inserts, the first row contains the experimental potency values of each analog and the second row the mean FW-predicted potency values (with the corresponding number of FW NBHs in parentheses). The third row contains the mean SVR-predicted potency values (with the corresponding number of individual prediction trials in parentheses).

AS: Analog series; EA: Existing analog; FW: Free-Wilson; ID: Identification; NBH: Neighborhood; SVR: Support vector regression.

Figure 6. Potency predictions for Free-Wilson virtual analogs.

(A) Box plots compare experimental potency value distributions of 11 ASs according to Figure 5A with potency predictions of corresponding FW VA populations. The y-axis reports logarithmic potency values and the x-axis the number of FW VAs per series. Numbers in parentheses are AS IDs. The experimental potency distribution of all EAs per series is displayed in light green, the FW-predicted VA potency distribution in red and the corresponding SVR-predicted distribution in orange. (B) Exemplary VAs (middle and right) are shown that were predicted to have higher potency than the most potent EA (left) of an AS active against the P2X purinoceptor 3 (AS 19; ChEMBL target ID 2998). In beeswarm plots below (color-coded according to the box plots), the exemplary compounds are indicated using arrows.

AS: Analog series; EA: Existing analog; FW: Free-Wilson; ID: Identification; NBH: Neighborhood; SVR: Support vector regression; VA: Virtual analog.

Figure 7. Free-Wilson neighborhood saturation scores.

N scores are shown for ASs yielding predictive models as a function of increasing FW EA fraction, defined as the proportion of FW EAs among all EAs. Dots represent ASs that are scaled in size by the number of analogs per series and color-coded according to the number of algorithmically generated FW VAs per series.

AS: Analog series; EA: Existing analog; FW: Free-Wilson; VA: Virtual analog.