Building structure activity relationships to avoid toxicity due to unwanted central nervous system ion channel activity Authors: Louisa A K Zolkiewski, Kimberly L Rockley, Ruth A Roberts, Hannah G Jennings, Magali-Anne Maizières, Karen Jones, James L McDonagh, Katie Newman, Ed J Griffen, Jessica E Stacey, Michael J Morton Published: 2026 ABSTRACT AbstractWe previously described an integrated in vitro liability assay for seizure, a new approach methodology (NAM) to reduce toxicity due to central nervous system (CNS) liability in drug discovery and development. Here we report the development of structure activity relationships (SAR) to guide drug design away from this liability. SAR test compounds were selected from the Enamine REadily AccesibLe (REAL) database using pharmacophore features and similarity to previous test compounds (amoxapine, diphenhydramine, quetiapine, 4-AP, linopirdine) or to ion channel positive reference compounds (bepridil, NS1619, quinidine, verapamil, XE991). These 88 compounds (10 parent compounds and 78 structurally related derivatives) were screened by automated electrophysiology in cell lines expressing human KV2.1, NaV1.2, the α1β2γ2 GABAA or α4β2 nicotinic receptors to generate IC50 values. Across all 4 ion channels, the derivative compounds exhibited increased, equivalent or reduced potency compared with the parent compounds, providing a complex and rich dataset for SAR. Regarding individual pharmacophoric features, statistical analysis identified 12 features significantly associated with activity at the α4β2 nicotinic receptor; 4 of these were also significantly associated with activity at KV2.1. Selected parent compounds (amoxapine, diphenhydramine, quetiapine) and structural analogues were screened for seizure-like activity in human induced pluripotent stem cell neurons using microelectrode array. The seizure-like phenotype was altered with the derivatives, as expected from the respective ion channel IC50 values. These data provide insight into specific substructures associated with seizure-like drug toxicity, offering the opportunity to avoid CNS liability in the development of novel compounds, saving time, money, and resources.Keywords: SAR, CNS toxicity, drug development, drug discovery, seizure, ion channels, MEA, hiPSC neurons, iLASeizure Central nervous system (CNS) toxicities, specifically drug-induced seizures, frequently occur during drug discovery and development (Cook et al. 2014; Authier et al. 2016), leading to project delays and failures. Compounds associated with CNS liability span a variety of therapy areas including CNS, cardiovascular, gastrointestinal, respiratory, and inflammation (Easter et al. 2009; Authier et al. 2016; Walker et al. 2018). Seizures are usually detected during preclinical rodent and nonrodent studies or even later, once the drug has entered clinical development, leading to a significant waste of time and resources. Progress has been made in seizure detection using automated video systems that record and analyze animal movements (Yip et al. 2019), and substantial effort has been invested in developing screening methods that could be used earlier in drug discovery, such as hippocampal brain slice (Easter et al. 2007) or zebrafish larval locomotor assays (Winter et al. 2008). However, these methods have had limited success, with concerns regarding throughput and translation of rodent and zebrafish data to humans. We have previously described an in vitro liability assay for seizure (iLASeizure) based on measuring electrical activity in induced human pluripotent stem cell (hiPSC) neurons using microelectrode array (MEA), coupled with a neuronal ion channel panel (Rockley et al. 2023). Activity of 15 known seizurogenic and 1 negative compounds at 15 seizure-related ion channels (NaV1.1, NaV1.2, NaV1.6, KV7.2/7.3, KV7.3/7.5, KV1.1, KV4.2, KCa4.1, KV2.1, KV3.1, KCa1.1, CaV2.1, and the α1β2γ2 GABAA, α4β2 nicotinic and GluN1/2A NMDA receptors) was assessed using automated patch-clamp electrophysiology. The data demonstrated inhibition of function across ion channels with some notable patterns. Specifically, there was a similar inhibition profile across the 3 sodium (NaV) channels, the KV2.1 potassium channel was inhibited the most among the CNS active therapies and the nicotinic α4β2 receptor was the most frequently hit channel (Rockley et al. 2023). A further 7 negative compounds have been assessed using iLASeizure since the publication of this work. Of these 23 screened compounds, there were 13 true positives, 7 true negatives, 1 false positive, and 2 false negatives. This resulted in predictivity, sensitivity, specificity, and positive and negative predictive values for the assay of 87%, 87%, 88%, 93%, and 78%, respectively. For compounds identified as a potential seizure risk, medicinal chemists face the challenge of redesigning the molecule to avoid the potential termination of a promising lead series. The ability to link chemical structures to molecular properties, such as potency or stability, is known as structure activity relationship (SAR); for example, “adding a chlorine at 2-position adds a 3-fold potency but reduces solubility 10-fold” (Griffen et al. 2011). Methods to establish these relationships and develop rules to guide design have been used for conventional small molecules, novel modalities, and even materials design (Allen et al. 2024; McDonagh et al. 2024). As seen with the human Ether-à-go-go-Related Gene (hERG), a potassium channel associated with cardiac liability, understanding, and modelling SAR can direct redesign to avoid substructures associated with unwanted hERG activity (Gund 1977; Garrido et al. 2020; Creanza et al. 2021). Alternatively, considering the impact of specific changes, such as “How much potency does adding a chlorine at the 2-position generally give?” can prove informative. Indeed, using quantitative SAR (QSAR) approaches, several well-established pharmacophore features have been associated with hERG inhibition, which can be used to guide strategies to design out hERG liabilities (Garrido et al. 2020). A pharmacophore is defined as “a set of structural features in a molecule that is recognized at a receptor site and is responsible for the molecule’s biological activity”; typical features include hydrophobic centroids, aromatic rings, hydrogen bond acceptors or donors, cations, and anions (Gund 1977; Wermuth et al. 1998). Inspired by the hERG successes, we embarked on an SAR study to identify the pharmacophore features responsible for the inhibition of seizure-related ion channels: Nav1.2, Kv2.1, and the α1β2γ2 GABAA and α4β2 nicotinic receptors. For this study, the key pharmacophore features of selected original test compounds (amoxapine, diphenhydramine, quetiapine, 4-AP [4-Aminopyridine], linopirdine) plus several ion channel positive reference compounds (bepridil, NS1619, quinidine, verapamil, XE991) were identified. Structural analogues were extracted from the Enamine REadily AccesibLe (REAL) database by scaffold searching and then selected based on the similarity of the pharmacophore features to the parent compounds. The Enamine REAL database is an extremely large enumerated database of synthetically feasible molecules (>1.2 billion) that can assist in SAR investigations (Grygorenko et al. 2020; “REAL Database—Enamine,” n.d.). We also investigated electrical activity by MEA in hiPSC neurons to confirm that significant differences in ion channel potencies translated to a phenotypic change between parent and derivative compounds. The data demonstrate that certain pharmacophore features present a significant risk of CNS ion channel inhibition, which translates to perturbed electrical signaling in hiPSC neurons. Together, this highlights the significant value of developing a robust predictive model that can assess SAR of compounds with seizurogenic potential to enable structural chemists to design-out CNS liability in the development of novel compounds by ensuring test molecules do not have high-risk groups, thus eliminating costly project failures. Materials and methods Ion channels Chinese hamster ovary (CHO-K1) or human embryonic kidney (HEK293) cells stably expressing the wild-type human ion channels Nav1.2, Kv2.1 and the α1β2γ2 GABAA and α4β2 nicotinic receptors were generated by standard recombinant techniques (Green and Sambrook 2012). The voltage-gated sodium and potassium channels were created at ApconiX, and the ligand-gated channels were purchased from B’SYS GmbH (Witterswil, Switzerland). Cell culture CHO-K1 cells were cultured in F-12 Ham medium (cat. N6658, Sigma-Aldrich, Merck, St. Louis, MO, United States) supplemented with 10% v/v fetal bovine serum (cat. SV30160.03, Cytiva, Marlborough, MA, United States), 1% v/v nonessential amino acids (cat. 11140-050, Gibco, ThermoFisher Scientific, Waltham, MA, United States), 1% v/v penicillin-streptomycin (cat. P4333, Sigma-Aldrich), and the appropriate selection antibiotics. HEK293 cells were cultured in a 50:50 mix of F-12 Ham and Dulbecco’s Modified Eagles Medium (cat. SH30081.01, Cytiva) supplemented with 10% v/v fetal bovine serum, 1% v/v nonessential amino acids, 1% v/v penicillin-streptomycin, and the appropriate selection antibiotics. Compounds The compound set used to investigate SAR at CNS ion channels consisted of 10 parent compounds, either seizure-associated compounds from the initial in vitro assay validation study (Rockley et al. 2023) or ion channel positive reference compounds, purchased from either Sigma-Aldrich or TOCRIS (Bristol, United Kingdom). A further 78 compounds were then selected and purchased from the Enamine REAL database (Kyiv, Ukraine). Scaffold searches were performed for each of the reference compounds within the Enamine REAL database, the closest 140 compounds in each case by extended-connectivity fingerprints (ECFP) were then further filtered by counts of pharmacophore features. Pharmacophore features were defined as a pair of substructures connected by their shortest topological path, similar to that described by Reutlinger et al. (2013). The shortest path distance between 2 features is within the same group or a direct link, and the longest link considered was 8 atoms separated. A diagrammatic example is provided in Fig. S1. The pharmacophore features were hydrophobic group; hydrogen bond donors; hydrogen bond acceptors; aromatic ring attached; aliphatic ring attached; halogen bonds; basic group; acidic group (Table 1). Compounds were then picked to iteratively remove or reduce the number of each pharmacophoric feature present in the reference compound and the candidate compound to represent a test of each feature. This resulted in parent and (derivative test compound numbers) as follows: Amoxapine (n = 10), diphenhydramine (n = 9), quetiapine (n = 10), 4-AP (n = 10), linopirdine (n = 8), bepridil (n = 9), NS1619 (n = 9), quinidine (n = 9), verapamil (n = 2), and XE991 (n = 2). Table 1.Pharmacophore descriptions used to define a pharmacophore feature.PharmacophoreAbbreviationDescriptionBasic GroupBGAtom or group most likely protonated at pH 7.4 (definitions derived from MedChemica proprietary SMARTS library)Acidic GroupAGAtom or group most likely deprotonated at pH 7.4, includes N and C acids (definitions derived from MedChemica proprietary SMARTS library)Hydrogen Bond AcceptorHADefinitions derived from Taylor et al. (2012)Hydrogen Bond DonorHDDefinitions derived from Taylor et al. (2012)HydrophobicLHC4 or greater, cyclic or acyclic alkyl groupAromatic AttachmentARConnection of any group to an aromatic atom excluding connections within ringsAliphatic AttachmentRRConnection of any atom to an aliphatic group not in a ringHalogen bondsHalF, Cl, Br, I, CF3, OCF3A pharmacophore features is a pair of pharmacophores connected by their shortest topological path. The shortest path distance between 2 distances are: Same group, direct link and 1 to 8 atoms. Pharmacophore descriptions used to define a pharmacophore feature. A pharmacophore features is a pair of pharmacophores connected by their shortest topological path. The shortest path distance between 2 distances are: Same group, direct link and 1 to 8 atoms. Stock solutions were made in dimethyl sulfoxide (DMSO; cat. D8148, Sigma-Aldrich) to 33 mM and stored at −20°C until use. Screening of compounds QPatch Automated patch-clamp electrophysiology experiments (QPatch II, Sophion Biosciences, Ballerup, Denmark) (Bell and Dallas 2018) were performed as previously described (Rockley et al. 2023). Briefly, stock compounds in DMSO were diluted to their top test concentration in HEPES-buffered saline. Single-cell ionic currents were measured in whole-cell configuration at room temperature (25°C). Compounds were incubated for at least 120 s. Concentration-response curves were generated by cumulative addition of compound with concentrations low to high. In all cases, steady-state inhibition was achieved before the next concentration of compound was added. The difference in current between the pre- and post-compound measurements was exported and analyzed using GraphPad Prism software (version 10.6.1; GraphPad Software, Boston, MA, United States). A minimum of 6- (KV2.1, NaV1.2), 5- (α1β2γ2 GABAA receptor), or 4-point (α4β2 nicotinic receptor) concentration-response curves were generated for each compound to generate IC50 (half-maximal inhibitory concentration) values. IC50 values were obtained from a 4-parameter logistic fit of normalized concentration-response data. Data were transformed to the pIC50 (the negative logarithmic measure of the IC50), a value often used in pharmacology to easily compare compound potency, where a higher value indicates greater potency. Microelectrode array iCell GlutaNeurons (cat. R1034, FUJIFILM Cellular Dynamics Inc. [CDI], Maddison, WI, United States) and iCell astrocytes (cat. R1092, FUJIFILM CDI) were mixed and seeded into 48-well MEA plates (16 electrodes/well; Axion Biosystems, Atlanta, GA, United States) precoated with 0.07% v/v Polyethylenimine (cat. 181978, Sigma-Aldrich) at a final ratio of 8:3 (GlutaNeurons: Astrocytes), following manufacturer’s protocols. The lot numbers and final cell densities per well for each biological replicate are provided in Table 2. Cells were cultured in complete BrainPhys™ media: BrainPhys™ media (cat. 05790, Stem Cell Technologies, Vancouver, Canada) supplemented with 1% v/v iCell Nervous System Supplement (cat. M1031, Fujifilm CDI), 2% v/v iCell Neural Supplement B (cat. M1029, Fujifilm CDI), 1% v/v N2 supplement (cat. 17502048, ThermoFisher Scientific), 1% v/v penicillin-streptomycin, and 1 µg/mL laminin (cat. 23017-015, Gibco) with 50% media changes performed every 24 to 72 h. Spontaneous electrical activity was assessed weekly during cell maturation on day in vitro (DIV) 5, 12, and 19. Compounds were added on—DIV17-30. Table 2.Cell lots used in hiPSC neuronal MEA assay.MEA replicateGlutaNeuron lot numberAstrocyte lot numberFinal cell density per well1107487 (81% glutamatergic neurons/19% GABAergic neurons)10758097,200 glutamatergic neurons, 22,800 GABAergic neurons and 20,000 astrocytes2108288 (84% glutamatergic neurons/16% GABAergic neurons)107638100,800 glutamatergic neurons, 19,200 GABAergic neurons and 20,000 astrocytes3107729 (83% glutamatergic neurons/17% GABAergic neurons)10758099,600 glutamatergic neurons, 20,400 GABAergic neurons and 20,000 astrocytesiCell GlutaNeuron and iCell Astrocyte lot numbers used for 3 independent biological replicates of the MEA analysis experiment. Final seeding densities of each cell type, per well, is also provided. Cell lots used in hiPSC neuronal MEA assay. iCell GlutaNeuron and iCell Astrocyte lot numbers used for 3 independent biological replicates of the MEA analysis experiment. Final seeding densities of each cell type, per well, is also provided. Stock compounds were solubilized at 1000× their top concentration in DMSO. Serial dilutions in complete BrainPhys™ media were prepared at 10x final concentration which were added directly to media in the wells. A 0.1% DMSO vehicle control was included on every plate. Before compound addition, the MEA plate was equilibrated in the MEA Pro instrument (Axion Biosystems) for at least 15 min. Then, 15-min baseline recordings were taken, and 15-min treatment recordings were taken after 1 h incubation with test compounds. The sampling frequency was 12.5 kHz. The digital low-pass filter was 3 kHz and the high-pass filter was 200 Hz. Spike files were created from the original 15-min recordings for analysis. Raster plots of the first 150 s of the recordings were created using the Axion Neural Metric Tool (v4.1.1), and the effects of compound treatment on selected MEA parameters were determined using the Axion Axis Metric Plotting Tool (v2.5.1). The effect of each drug was measured relative to its own pre-drug baseline (from the same well). Compounds (n = 4) or vehicle (n = 6 to 12) were applied directly to wells, and the average change from baseline was calculated. The DMSO vehicle control is presented for each parameter and usually shows minimal changes (±20%). A 1-way ANOVA with Šídák’s multiple comparisons test was performed on parent and derivative compound data (Graphpad Prism software v10.6.1); P < 0.05 was considered significant. Many different MEA parameters can be used to describe electrical activity, such as weighted mean firing rate (WMFR), interspike interval (ISI) coefficient of variation, area under normalized cross correlation (AUNCC), network burst frequency, network burst duration, and number of spikes per network burst (see Table 3 for parameter definitions). Similar parameters have been selected in analyses by other groups (Kreir et al. 2018; Tukker et al. 2020a, 2020b). Table 3.Definitions of selected MEA parameters, as defined by Axion Biosystems.Firing rate parametersDefinitionWeighted mean firing rate (Hz)The mean firing rate based on only electrodes with activity greater than minimum spike rate.ISI coefficient of variationThe coefficient of variation (SD/mean) of the interspike interval, the time between spikes. This is a measure of spike regularity.Network burst parametersNetwork burst frequency (Hz)Total number of network bursts divided by the recording time.Network burst duration (s)Average time from the first spike to the last spike in a network burst.Number of spikes per network burstAverage number of spikes in a network burst.Synchrony parameterArea under normalized cross-correlationArea under well-wide pooled interelectrode cross-correlation normalized to the auto-correlations. High values indicate greater synchronicity of the network. Definitions of selected MEA parameters, as defined by Axion Biosystems. Pharmacophore analysis Using the 2-sided Mann–Whitney U-test as implemented in Scipy (version 1.16.3), P-values were calculated using the Bonferroni Correction (BC) to account for multiple comparisons. Strict criteria were applied to ensure that pharmacophore features identified as significant were those of the greatest risk and not statistical artifacts. First, a feature was only considered significant if the Mann–Whitney U-test P-value was less than 0.05/207 = 0.00024 (BC corrected 95% threshold) when comparing the shift in average pIC50 value between the group of compounds with or without the pharmacophore feature. Theoretically, there are (N[N 1]) x 10/2 = 360 pharmacophore features within the compound set tested, where N = 8 (i.e. the number of pharmacophore features), 10 is the number of distances considered between features and the division by 2 is due to the symmetry of the dyad pairs. Here, only 207 pharmacophore features were actually observed in the molecules screened, hence why this was used for the BC. Secondly, there needed to be a minimum difference of 0.3 in the pIC50 value between the mean of compounds with and without a specific pharmacophore feature. This delta pIC50 (ΔpIC50) threshold of 0.3 corresponds to a 2-fold change in IC50 values and is commonly used within QSAR modelling as a threshold to define activity cliffs and account for assay variability (Cramer 2011; Shen et al. 2024). Finally, 8 or more examples of compounds with or without a given feature were required in the dataset tested. Given the small size of the dataset, this value was chosen to increase the likelihood that any statistically significant data were not random. For example, when conducting a binomial test, the probability of observing 8/8 positive measurements at random (based on a ΔpIC50 of 0.3) = 0.007812, which is below the P < 0.05 threshold required to reject a null hypothesis. Graphical representation of structural and spatial distribution of compounds To visualize the distribution of IC50 values across the compound set, each parent molecule and the associated structural relatives from the Enamine REAL set were extracted and grouped by IC50 values at each of the 4 ion channels screened. A boxplot was constructed to represent the distribution of IC50 values for each group. The parent molecule’s IC50 value was highlighted in blue and each associated molecule’s IC50 was overlayed as a black dot. Visualizations of the structures of the Enamine REAL molecules relative to their respective parent molecules were generated using Networkx (v3.5). These images show the parent molecule in the center and each associated molecule around the outside connected to the parent with a molecular similarity (ECFP4 [2048 bits] and Tanimoto similarity) to summarize the degree of similarity between the two. The IC50 values and the difference (ΔIC50) compared with the parent are quoted for each associated molecule. Molecule images were all rendered using the RDKit (v2023.9.6) (“RDKit: Open-Source Cheminformatics Software,” n.d.). Results Ion channel potency The potencies of the Enamine REAL compounds compared with the parent compounds are presented as representative graphs with IC50 values in Fig. 1 and summarized in Table 4. All raw IC50 values and pharmacophore feature labels for each compound are provided in Supplementary Data. At the KV2.1 ion channel, the potency of the parent compounds ranged from <2 to >100 μM. Eight of the 10 parent compounds were highly potent (IC50 <30 μM), with the exception of quetiapine and 4-AP. Against this background, 36 compounds exhibited similar potencies compared with the respective parent molecules, and 32 and 10 compounds had reduced and increased potencies, respectively. Notably, all the 4-AP derivative compounds remained inactive, in line with the parent molecule. Further, all the derivative compounds for NS1619, quinidine and verapamil had reduced potencies compared with their parent molecules, whereas all the XE991 derivative compounds were more potent than the parent (Table 4). Distribution of potencies of selected parent and derivative compounds across 4 seizure-related ion channels. Representative IC50 plots for amoxapine, bepridil, and diphenhydramine parent compounds (black circles) compared with derivative Enamine REAL compounds with the highest (red squares) and lowest (green triangles) potencies screened against seizure-related A. KV2.1, B. Nav1.2, C. α1β2γ2 GABAA and D. α4β2 nicotinic ion channels. Where Enamine REAL compounds with higher or lower potency were not identified during screening, alternative representative compounds have been shown. Data shown as mean ± SD from a minimum n = 3 cells. Table 4.Potency of Enamine REAL compounds compared with 10 parent compounds screened against CNS-related ion channels KV2.1, NaV1.2 and the α1β2γ2 GABAA and α4β2 nicotinic receptors.Ion channelParent compound4-APAmoxapineBepridilDiphenhydramineLinopirdineNS1619QuetiapineQuinidineVerapamilXE991KV2.1Parent compound IC50>100 μM6.7 μM2.4 μM5.5 μM8.8 μM7.6 μM31.4 μM1.2 μM1.5 μM19.2 μMIC50 <10 μM0777301002IC50 10–30 μM0302217000IC50 30–100 μM0010112210IC50 >100 μM10010270710NaV1.2Parent compound IC50>100 μM26.2 μM9.4 μM>100 μM>100 μM56.9 μM>100 μM>100 μM35.6 μM>100 μMIC50 <10 μM0000000000IC50 10–30 μM0101000000IC50 30–100 μM1432142001IC50 >100 μM9566758921α1β2γ2 GABAAParent compound IC50>100 μM2.6 μM>100 μM>100 μM>100 μM>100 μM>100 μM>100 μM>100 μM>100 μMIC50 <10 μM0200000000IC50 10–30 μM0200200100IC50 30–100 μM0100003100IC50 >100 μM10599697722α4β2 nicotinicParent compound IC5021.0 μM7.4 μM2.4 μM6.0 μM43.6 μM25.9 μM26.9 μM12.4 μM13.4 μM19.3 μMIC50 <10 μM0636100100IC50 10–30 μM1333319412IC50 30–100 μM2110431310IC50 >100 μM7020050100The number of Enamine REAL compounds with IC50 values of <10 μM, 10–30 μM, 30–100 μM and >100 μM relative to the respective parent compounds are included for each channel. Graphs showing IC50 values for the parent, and representative Enamine REAL derivative, compounds amoxapine, bepridil and diphenhydramine can be seen in Fig. 1 4-AP, 4-aminopyridine. Potency of Enamine REAL compounds compared with 10 parent compounds screened against CNS-related ion channels KV2.1, NaV1.2 and the α1β2γ2 GABAA and α4β2 nicotinic receptors. The number of Enamine REAL compounds with IC50 values of <10 μM, 10–30 μM, 30–100 μM and >100 μM relative to the respective parent compounds are included for each channel. Graphs showing IC50 values for the parent, and representative Enamine REAL derivative, compounds amoxapine, bepridil and diphenhydramine can be seen in Fig. 1 4-AP, 4-aminopyridine. For the NaV1.2 ion channel, the potency of the parent compounds ranged from <10 to >100 μM. Amoxapine and bepridil parent compounds were the most potent, with 6 of the remaining parent compounds inactive (IC50 >100 μM). Against this background, most of the compounds (45) had equivalent potency to the parent molecules, with 25 and 8 compounds with reduced or increased potency, respectively. Notably, all the quinidine derivative compounds remained inactive, in line with the parent molecule. In the case of bepridil and verapamil, all derivative compounds were less potent than their respective parent molecules (Table 4). At the α1β2γ2 GABAA receptor, the potency of the parent compounds ranged from <3 to >100 μM. The amoxapine parent compound was highly potent (IC50 = 2.6 μM), with all remaining parent compounds inactive (IC50 >100 μM). Against this background, most of the derivative compounds (63) had equivalent potency to the parent molecules, with 8 and 7 compounds with reduced or increased potency, respectively (Table 4). For the α4β2 nicotinic receptor, the potency of the parent compounds ranged from <3 to >40 μM. All 10 parent compounds were active, with at least moderate potency detected (IC50 <45 μM). Against this background, 37 compounds had equivalent potency to the parent molecules, and 36 and 5 compounds had reduced or increased potencies, respectively (Table 4). Data suitability for SAR The distribution of the 78 derivative compound IC50 values compared with the 10 parent molecules for each ion channel are shown in Fig. 2. These box plots illustrated variability in the data between ion channels, with KV2.1 and the α4β2 nicotinic receptor offering a complex and intriguing dataset. In comparison, at NaV1.2 and the α1β2γ2 GABAA receptor, the parent compounds themselves were often inactive and structural modifications usually resulted in no change, or loss of any potency. Distribution of potencies of parent and derivative compounds across 4 seizure-related ion channels. Distribution of IC50 values for Enamine REAL database compounds (black circles) relative to parent compounds (blue squares), separated by parent compound at seizure-related ion channels A. KV2.1, B. NaV1.2, C. α1β2γ2 GABAA receptor, and D. α4β2 nicotinic receptor. Mean values are represented by a dashed line, and median values are represented by a solid line. 4-AP, 4-aminopyridine. Figures S2–S5 show the structure and spatial distribution of IC50 values for the Enamine REAL database compounds relative to parent molecules at KV2.1, NaV1.2, and the α1β2γ2 GABAA and nicotinic-α4β2 receptors, respectively. Figure S6 shows the relative distribution of all 88 molecules in 2D using the t-SNE method. For KV2.1 and the α4β2 nicotinic receptor, there were some initial clusters of molecules which were consistent with IC50 classes (Fig. S6A and D). In contrast, for NaV1.2 and the α1β2γ2 GABAA receptor, most molecules fell into a single class hence little could be gained from these plots (Fig. S6B and C). A larger data set is required to explore this behaviour in more detail and apply similar methods at the other ion channels considered. Figure S7 shows boxplots of Tanimoto similarity between all molecules in the dataset. This provides a standard chemical similarity metric upon which to compare the molecules and can help to interpret the chemical space covered by the data. The mean and median similarities are typically low for all molecules as would be expected. Further, there are typically several similar molecules, this is not unexpected based on the parent molecule selection process. Given the dataset was relatively small, care was taken not to overinterpret these plots. However, small initial clusters emerging for 2 of the 4 ion channels is suggestive of regions of chemical spaces which are likely to exhibit higher or lower IC50 values at those targets. The range in IC50 values and spatial distribution across compounds at KV2.1 and the α4β2 nicotinic receptor presented a valuable dataset for SAR, featuring increasing, equivalent or decreasing potency with the structural modifications conducted. Significant pharmacophore features Tables 5 and 6 show significant pharmacophore features, each of which is defined as a 2D dyad pattern using proprietary SMARTS notations (“Daylight Theory: SMARTS—A Language for Describing Molecular Patterns,” n.d.), at KV2.1 and the α4β2 nicotinic receptor, respectively. There were 4 significant features at KV2.1: A hydrogen bond donor directly connected to an aromatic ring (HD_directlink_AR, ΔpIC50 = −0.77 [decreased potency]); a hydrogen bond donor separated from a hydrogen bond acceptor by a single atom (HD_1atom_HA, ΔpIC50 = −0.64 [decreased potency]); a hydrogen bond acceptor separated from a basic group by 2 atoms (HA_2atom_BG, ΔpIC50 = 0.52 [increased potency]) and an atom which is part of an aromatic system separated from a basic group by 4 atoms (AR_4atom_BG, ΔpIC50 = 0.51 [increased potency]) (Table 5). These 4 features were also significant at the α4β2 nicotinic receptor (HD_directlink_AR, ΔpIC50 = −0.53 [decreased potency]; HD_1atom_HA, ΔpIC50 = −0.54 [decreased potency]; HA_2atom_BG, ΔpIC50 = 0.56 [increased potency]; AR_4atom_BG, ΔpIC50 = 0.57 [increased potency]) and conferred the same direction of potency (Table 6). Eight additional pharmacophore features achieved statistical significance, with 7 of these associated with an increased potency and 1 associated with a decreased potency: A hydrogen bond donor separated from a halogen by 3 atoms (HD_3atom_Hal, ΔpIC50 = −0.56 [decreased potency]) (Table 6). In contrast, although there were some nominally differential features identified at the α1β2γ2 GABAA receptor and NaV1.2, none achieved statistical significance (Tables S3 and S4). Interestingly, an atom which is part of an aromatic system separated from a basic group by 4 atoms was also a differential feature at NaV1.2, showing increased potency as seen for KV2.1 and the α4β2 nicotinic receptor (Table S4). The full lists of the impact of differentially present pharmacophore features for all ion channels are shown in Tables S1–S4. Table 5.Statistically significant pharmacophore features at KV2.1.Pharmacophore featureMean pIC50 with featureWith feature (N =)Mean pIC50 without featureWithout feature (N =)ΔpIC50 P-valueHD_directlink_AR 4.127602 21 4.899596 67 −0.77199 4.77E-07 HD_1atom_HA 4.222548 20 4.860318 68 −0.63777 6.86E-05 HA_2atom_BG 5.020858 36 4.503878 52 0.51698 8.49E-05 AR_4atom_BG 5.019541 35 4.514502 53 0.505039 0.000125699 Data analyzed by 2-sided Mann–Whitney U-test with Bonferroni Correction; P < 0.0002415 considered significant (P = 0.05/207 [number of tests performed]). Positive (red text) and negative (green text) delta pIC50 (ΔpIC50) values represent increased and decreased potency, respectively, for compounds with the given pharmacophore feature. Bold shows pharmacophore features common across both KV2.1 and the α4β2 nicotinic receptor. Definitions and abbreviations for pharmacophores are listed in Table 1. The full dataset can be seen in Table S1. Statistically significant pharmacophore features at KV2.1. Data analyzed by 2-sided Mann–Whitney U-test with Bonferroni Correction; P < 0.0002415 considered significant (P = 0.05/207 [number of tests performed]). Positive (red text) and negative (green text) delta pIC50 (ΔpIC50) values represent increased and decreased potency, respectively, for compounds with the given pharmacophore feature. Bold shows pharmacophore features common across both KV2.1 and the α4β2 nicotinic receptor. Definitions and abbreviations for pharmacophores are listed in Table 1. The full dataset can be seen in Table S1. Table 6.Statistically significant pharmacophore features at the α4β2 nicotinic receptor.Pharmacophore featureMean pIC50 with featureWith feature (N =)Mean pIC50 without featureWithout feature (N =)ΔpIC50 P-valueHA_samegroup_BG4.89039564.304497320.5858927.54E-10 AR_4atom_BG 5.022346 35 4.449502 53 0.572843 3.16E-09 HA_2atom_BG 5.010487 36 4.446696 52 0.56379 5.98E-09 HD_directlink_AR 4.274813 21 4.803502 67 −0.52869 1.15E-06 AR_5atom_BG4.947246364.490479520.4567672.02E-06HA_2atom_HA4.909389444.445286440.4641032.23E-06 HD_1atom_HA 4.262416 20 4.799374 68 −0.53696 2.31E-06 AR_1atom_BG4.932171374.492459510.4397126.96E-06HD_samegroup_BG4.964284314.521279570.4430052.43E-05BG_2atom_BG5.193379114.603618770.5897613.67E-05HD_3atom_Hal4.193404124.75374976 −0.560353.68E-05AR_2atom_BG4.882711374.528342510.3543690.000202Data analyzed by 2-sided Mann–Whitney U-test with Bonferroni Correction; P < 0.0002415 considered significant (P = 0.05/207 [number of tests performed]). Positive (red text) and negative (green text) delta pIC50 (ΔpIC50) values represent increased and decreased potency, respectively, for compounds with the given pharmacophore feature. Bold shows pharmacophore features common across both KV2.1 and the α4β2 nicotinic receptor. Definitions and abbreviations for pharmacophores are listed in Table 1. The full dataset can be seen in Table S2. Statistically significant pharmacophore features at the α4β2 nicotinic receptor. Data analyzed by 2-sided Mann–Whitney U-test with Bonferroni Correction; P < 0.0002415 considered significant (P = 0.05/207 [number of tests performed]). Positive (red text) and negative (green text) delta pIC50 (ΔpIC50) values represent increased and decreased potency, respectively, for compounds with the given pharmacophore feature. Bold shows pharmacophore features common across both KV2.1 and the α4β2 nicotinic receptor. Definitions and abbreviations for pharmacophores are listed in Table 1. The full dataset can be seen in Table S2. Figure 3 shows box plots of the impact of the 4 common pharmacophore features on pIC50 values at KV2.1 (Fig. 3A) and the α4β2 nicotinic receptor (Fig. 3B). For both KV2.1 and the α4β2 nicotinic receptor, the presence of either a hydrogen bond donor directly connected to an aromatic ring (panel 1) or hydrogen bond donor separated from a hydrogen bond acceptor by a single atom (panel 2) resulted in lower pIC50 values, equating to lower potency. In contract, the presence of a hydrogen bond acceptor separated from a basic group by 2 atoms (panel 3) or an atom which is part of an aromatic system separated from a basic group by 4 atoms (panel 4) resulted in higher pIC50 values, equating to higher potency. Generally, the pIC50 values were more variable at KV2.1 with these 4 features compared with without (Fig. 3A). In contrast, the pIC50 values were equally variable at the α4β2 nicotinic receptor with or without these 4 features (Fig. 3B). Impact of significant pharmacophore features at KV2.1 and the α4β2 nicotinic receptor. pIC50 values are shown for 88 compounds screened against A. KV2.1 or B. the α4β2 nicotinic receptor with or without the specified pharmacophore feature (see Tables 5 and 6 for details). An interesting question followed from this analysis: “do any of the significant pharmacophore features also impact other safety targets?”. We therefore additionally screened the 2 hERG datasets retrieved from ChEMBL (version 36). The molecules identified had measured Ki and/or EC50. The structures were cleaned by desalting, normalizing protonation states and standardizing tautomers. The Ki and EC50 values were converted to pKi and pEC50 as -log10(x) where x is the Ki or EC50 value. All statistical tests were applied to the transformed pKi and pEC50 values. These datasets were screened using the same featurization where pKi encountered 306 pharmacophore features and the pEC50 encountered 313 pharmacophore features out of the 360 possible features. The BC correction was applied based upon the number of encountered pharmacophore features. As with seizure, we split the datasets based on the molecules containing, or not, each pharmacophore feature independently and tested for a statistically significant shift in the average pKi or pEC50. We found that none of the 4 statistically significant pharmacophore features identified for seizure were identified for hERG. This suggested these identified features for seizure showed at least some specificity to those targets. The hERG datasets are provided in Supplementary Data. Figure S8 shows the pharmacophore features which were statistically significant at hERG for both pKi and pEC50. MEA analysis MEA analysis of amoxapine, diphenhydramine, quetiapine, and selected Enamine REAL derivate compounds was conducted to investigate if the modified potency at ion channels translated into changes in seizurogenic phenotypes at a whole-cell level. Amoxapine showed a clear seizurogenic phenotype compared with baseline activity at 3 μM (Fig. 4A, panel 1 raster plots), with a high potency at KV2.1, the α4β2 nicotinic and α1β2γ2 GABAA receptors, and moderate potency at NaV1.2 (Fig. 4A). The derivative molecule Z4456062749 showed a less potent seizurogenic phenotype (Fig. 4A, panel 2 raster plots) and was associated with lower potency at NaV1.2 and the α1β2γ2 GABAA receptor (Fig. 4A). However, this was not associated with statistically significant differences in MEA parameters compared with the parent compound (Fig. 4B). The derivative PV-002730229172 showed an even less potent phenotype (Fig. 4A, panel 3 raster plots) quantified by statistically significant differences in several MEA parameters (Fig. 4B). For example, PV-002730229172 exhibited significantly reduced WMFR (P = 0.0032) with an increased network burst duration (P = 0457). This was associated with considerably lower potencies at all 4 ion channels (Fig. 4A). The 3 compounds (Fig. 4C) share many structural similarities, with the loss of the halogen group (chlorine; Cl) from the amoxapine parent an obvious change as seizure risk is reduced from left to right. Noticeably, shifting the piperazine NH group position, through ring substitution to a diazepane (in compound Z4456062749), or replacing a hydrogen with a bulkier and flexible ethyl chain (in compound PV-002730229172) could alter the accessibility of an otherwise relatively rigid group, which may impact binding at the ion channels screened. MEA analysis of amoxapine and derivatives at 3 μM. A. Representative raster plots depicting 0 to 150 s activity at baseline (top row) and 60 min postdose (bottom row) with parent or derivative Enamine REAL compounds, Z4456062749 and PV-002730229172. IC50 values at 4 ion channels are shown below raster plots (black text = similar potency and green text = lower potency of Enamine REAL compound compared with parent at the given ion channel). B. Key MEA parameters highlighting significant differences. Data shown as mean ± SD from n = 3 biological replicates. Data analyzed by 1-way ANOVA with Šídák’s multiple comparisons test; *P < 0.05,**P < 0.01. C. Chemical structures of compounds show areas of structural similarity (green) and differences (purple). AUNCC, area under normalized cross-correlation; ISI, interspike interval; WMFR, weighted mean firing rate. Diphenhydramine showed a seizurogenic phenotype compared with baseline activity at 10 μM, and was associated with potent inhibition of KV2.1 and the α4β2 nicotinic receptor (Fig. 5A, panel 1). The derivative Z162576456 showed a reduced seizure phenotype in the MEA raster plots and had a similar ion channel profile with slightly reduced potency at KV2.1 and the α4β2 nicotinic receptor (Fig. 5A, panel 2). This difference was quantified by a statistically significant increase in ISI coefficient of variation (P = 0.0168) and number of spikes per network burst (P = 0.0011, Fig. 5B), in line with vehicle control values. In contrast, the derivative molecule Z1589066639 showed a clear seizure phenotype, resulting in a complete dysregulation of neuronal activity, associated with higher potency at KV2.1, NaV1.2 and the α4β2 nicotinic receptor (Fig. 5A, panel 3). This was reflected in significant decreases in several MEA parameters, including weighted mean firing rate (P = 0.0093), ISI coefficient of variation (P = 0.0199), network burst duration (P = 0.0343), number of spikes per network burst (P = 0.0159), and area under normalized cross correlation (P = 0.0115) compared with the parent compound (Fig. 5B). The chemical structures of the 3 compounds (Fig. 5C) show several areas of similarity and differences between Z162576456 and Z1589066639 compared with diphenhydramine. Replacing the benzene ring with a pyridine ring and replacing the ether (O) group with a secondary amine (NH) functionality from the parent molecule to Z162576456 may offer some reduction in seizure risk. Noticeably, addition of a halobenzyl ring and replacing the ether (O) group with a secondary amine in Z1589066639 could confer increased seizure risk. Clearly, the ether to amine occurs in both with opposing seizure risk outcomes. As some of these structural modifications are not independent, i.e. not the sole modification between 2 of the structures, we cannot assign the seizurogenic change to 1 of these modifications, as there are potentially joint effects between the modifications. Further work is needed to establish these links. MEA analysis of diphenhydramine and derivatives at 10 μM. A. Representative raster plots depicting 0 to 150 s activity at baseline (top row) and 60 min postdose (bottom row) with parent or derivative Enamine REAL compounds, Z162576456 and Z1589066639. IC50 values at 4 ion channels are shown below raster plots (black text = similar potency, green text = lower potency, and red text = higher potency of Enamine REAL compound compared with parent at the given ion channel). B. Key MEA parameters highlighting significant differences. Data shown as mean ± SD from n = 3 biological replicates. Data analyzed by 1-way ANOVA with Šídák’s multiple comparisons test; *P < 0.05, **P < 0.01. C. Chemical structures of compounds show areas of structural similarity (green) and differences (purple). AUNCC, area under normalized cross-correlation; ISI, interspike interval; WMFR, weighted mean firing rate. Quetiapine showed a mild seizurogenic phenotype compared with baseline activity at 10 μM, and showed moderate inhibition of KV2.1 and the α4β2 nicotinic receptor (Fig. 6A, panel 1). The derivative Z3649159890 showed a slightly less seizurogenic phenotype and slightly reduced potency at the α4β2 nicotinic receptor (Fig. 6A, panel 2). However, this was not associated with statistically significant differences in MEA parameters compared with the parent compound (Fig. 6B). In contrast, the derivative molecule PV-004814057247 showed a potent seizurogenic phenotype, associated with a higher potency at all 4 ion channels (Fig. 6A, panel 3). There were significant changes in several MEA parameters including an increased network burst frequency (P = 0.0014), and decreased ISI coefficient of variation (P = 0.0398), network burst duration (P = 0.0011), and number of spikes per burst (P = 0.0207) compared with the parent compound (Fig. 6B). The chemical structures of the 3 compounds (Fig. 6C) revealed significant areas of difference; the major structural differences occurred on the substituents of the piperazine ring. The parent molecule possessed an extended bifunctional chain containing an ether and primary alcohol. The 2 sampled molecules, Z3649159890 and PV-004814057247, respectively possessed a cyclized version of the parents chain based on a tetrahydrofyran ring and a bifurcated monofunctional chain which retains the ether functionality of the parent molecule. Additionally, both of the sampled molecules substituted the main Dibenzo[B, F][1,4]thiazepine framework for the Dibenz(B, F)(1,4)Oxazepine framework removing the sulfur and replacing with an oxygen. The 2 sampled molecules showed divergent seizurogenic potential suggesting the framework modification leads to minimal impact. The pyrazine substituent modifications which differ between the 2 sampled molecules are more likely to be the source of this divergent behaviour, with the bulkier ring system (Z3649159890) modification raising the IC50 at 2 of the ion channels and the bifurcated chain (PV-004814057247) reducing the IC50 at all of the ion channels screened. MEA analysis of quetiapine and derivatives at 10 μM. A. Representative raster plots depicting 0 to 150 s activity at baseline (top row) and 60 min postdose (bottom row) with parent or derivative Enamine REAL compounds, Z3649159890 and PV-004814057247. IC50 values at 4 ion channels are shown below raster plots (black text = similar potency, green text = lower potency, and red text = higher potency of Enamine REAL compound compared with parent at the given ion channel). B. Key MEA parameters highlighting significant differences. Data shown as mean ± SD from n = 3 biological replicates. Data analyzed by 1-way ANOVA with Šídák’s multiple comparisons test; *P < 0.05,**P < 0.01. C. Chemical structures of compounds show areas of structural similarity (green) and differences (purple). AUNCC, area under normalized cross-correlation; ISI, interspike interval; WMFR, weighted mean firing rate. Discussion CNS liability, specifically drug-induced seizures, is a leading cause of project delays and failures during drug discovery and development. CNS toxicity is associated with compounds spanning therapy areas including CNS, cardiovascular, gastrointestinal, respiratory, and inflammation (Easter et al. 2009; Cook et al. 2014; Walker et al. 2018). Typically, CNS liabilities are not assessed until late-stage preclinical in vivo studies, often going undetected until clinical phase studies, after significant investment of time and money. We have previously shown that a panel of CNS ion channels can be used to screen for toxicity due to seizure liability in lead generation/lead optimization and can provide mechanistic insight where seizures have been observed in GLP toxicology studies or in clinical development (Rockley et al. 2023). Here, to assist in the development of therapies without liability, we aimed to develop SAR by identifying the specific structural characteristics of compounds associated with ion channel-related seizurogenic activity. As a proof-of-concept study, we initially selected 4 ion channels (from the 15 ion channels in the iLASeizure panel) that would be representative of the key ion channel families associated with seizure risk. Compounds were selected from the Enamine REAL database using pharmacophore features to assess similarity to reference compounds. The final set consisted of 10 parent and 78 derivative compounds. We initially assessed ion channel activity by automated patch-clamp as this is a relatively high-throughput assay and thus more practical than MEA for the >350 datapoints that were to be evaluated. The pharmacophore analysis of the α4β2 nicotinic receptor provided the richest dataset, with 63 differential features detected (where pIC50 values differ with the pharmacophore feature vs without), 12 of which reached statistical significance. Within this dataset, several features conferred decreased potency but most of the significant features increased potency, which was unexpected. The pharmacophore analysis of KV2.1 yielded 37 differential features, 4 of which achieved statistical significance. Notably, these 4 features were also significantly different at the α4β2 nicotinic receptor. In both ion channels, 2 pharmacophore features increased potency whereas 2 features decreased potency. The SAR analysis for NaV1.2 and the α1β2γ2 GABAA receptor revealed many differential pharmacophore features, but none achieved statistical significance. Interestingly, 1 differential pharmacophore feature at NaV1.2, showed increased potency in line with KV2.1 and the α4β2 nicotinic receptor. The commonality of these preliminary signals between ion channels highlights the value in developing SAR for CNS-related toxicities. This warrants further investigation to build a larger dataset that can be utilized to build a robust predictive CNS SAR model. Typically, CNS-penetrating molecules are of modest molecular weight, low polar surface area, have pKa ∼7.5 to 10.5 and possess few hydrogen bond donors (Pajouhesh and Lenz 2005). The identified pharmacophore features provide defined chemical substructures: Although considering physicochemical properties of these substructures in isolation is unlikely to be beneficial (as the molecular scale determines CNS penetration not isolated substructural elements), we note some links. For example, the pharmacophores leading to increased potency both include bases. Although none of the identified significant pharmacophores contain acidic groups which would be excluded by the typical CNS-penetrating molecular pKa range. Further, consideration should be given to the effect of metabolites. Changes in molecular structure due to metabolism inevitably modify a molecule’s properties and behaviour, but these changes remain challenging to predict. For example, if a primary metabolite has increased seizure potential and it crosses the blood–brain barrier, then this may induce seizures. However, usually, metabolism leads to more polar compounds that are easier to excrete and are less CNS-permeable, therefore are less likely to be seizurogenic in vivo. It is likely that additional SAR will be introduced by primary metabolites, warranting consideration in future work. For machine learning projects, it is common to develop applicability domains to approximately define where a model can reasonably be applied. A question that emerges for this work is “can we define a suitable applicability domain?”. Given the size of the dataset for this work the standard feature structure-based spaces are quite sparse and descriptor/property spaces show very limited ranges in some dimensions which can make an applicability domain definition less reliable. This work provides a preliminary analysis and small dataset in the supporting information. This enables others to build from this work and combine the data with new measurements to circumvent some of these issues. As the ion channel data indicated SAR, we selected representative molecules for MEA (a lower throughput technology) to confirm that the altered ion channel potency for derivative molecules translated into phenotypic differences in seizure in hiPSC neurons. The data showed that altered IC50 values translated to differential changes in MEA for the 3 parent molecules tested compared with their derivatives. A chemical structure analysis revealed areas of similarity and difference associated with these potency changes, guiding structural chemistry away from unwanted structural features. There were some notable observations for amoxapine in that shifting the piperazine NH group position, through ring substitution to a diazepane, or replacing a hydrogen with a bulkier and flexible ethyl chain altered potency; this might be due to altering the accessibility of an otherwise relatively rigid group, which could affect binding at the ion channels screened. Similarly for diphenhydramine, replacing the benzene ring with a pyridine ring and replacing the ether group with a secondary amine functionality appeared to confer reduced seizure risk, whereas addition of a halobenzyl ring, along with the secondary amine substitution, increased seizure risk. Unfortunately, in the current dataset, from the diphenhydramine parent, the secondary amine replacement occurs jointly with other replacements and with divergent IC50 impact, so we cannot decompose the individual group contributions to seizurogenic risk mitigation any further at this stage. Secondary safety pharmacology data are included in Investigational New Drug applications, yet only hERG screening requirements are explicitly described in ICH S7B guidelines (ICH 2005). Recently, there have been calls to revise and update the ICH S7A guidelines following their original publication in 2000 (Bowes et al. 2012; Lynch et al. 2017; Valentin and Leishman 2023, 2025; Brennan et al. 2024). A key proposal for the revisions focuses on human-based in vitro and in silico secondary pharmacology models. Indeed, Brennan et al. noted that the field could be revolutionized with the combination of high-quality functional data and machine learning models (Brennan et al. 2024). Recently, there has been a significant global commitment to reduce reliance on animal studies across the life sciences, particularly in preclinical safety testing (FDA 2022, 2025; UK Government 2025). Focus has been on developing and implementing NAMs (in vitro systems, in silico models, ex vivo tissues) for regulatory decision-making to increase accuracy and ethical acceptance (FDA 2025). Reflecting this commitment, the iLASeizure work was cited by The US Food and Drug Administration (US FDA) in their publication on promising new approach methodologies (NAMs) (Avila et al. 2023). Correspondingly, between 2015 and 2024 there were considerable increases of 35% and 27% in the use of stem cell-derived neurons and neuronal/glial co-cultures, respectively, for seizure liability testing within the safety pharmacology industry (Authier et al. 2016, 2025). This centrally positions iLASeizure and the proposed CNS SAR safety model as a combined human-based approach that could be implemented earlier in the drug discovery process to assess safety in line with these proposals. Since the 1960s, QSAR approaches have been widely used in drug discovery to predict blood–brain barrier penetration and pharmacological activity based on chemical substructures (Gupta 1989). Today, several SAR platforms are commercially available from Certara (D360™ (Certara, n.d.) and Libra™ (“Libra,” n.d.)), BioAscent (BioAscent, n.d.), and Solvonis (“Initiation of artificial intelligence-supported CNS drug discovery programme—Solvonis Therapeutics,” n.d.). However, the software and data used to build these platforms are all proprietary. Further, while some have a CNS focus, their primary function is to predict brain penetrance and primary pharmacology for drug candidate prioritization. Prediction models for drug-induced seizure liability have been published; however, these are not without caveats. For example, while a support vector machine classification prediction model with an “excellent” prediction accuracy of 86% has been published (Zhang et al. 2011), the model lacked mechanistic or target prediction insights and was limited to general predictivity. More recently, publicly-available ChEMBL data from 230 (196 seizure-positive and 34 seizure-negative) compounds was used to build 3 machine learning models to predict seizure-liability across 67 CNS targets (Lane et al. 2025). Initially, these models appeared to correctly predict the activity and seizure-liability of the molecules assessed; however, the prediction accuracy was not quantified. Further, there were concerns over the quality and availability of the data used to train the models, with the authors noting that the models predicted moderate activity of eglumetad at the NMDA receptor, whereas experimental testing revealed no activity (no binding at >100 μM) (Lane et al. 2025). Finally, while these models have been shown to predict drug-induced seizure liability across several CNS targets, there is no report of which substructures are associated with an increased seizure risk. In contrast, the data we present here, and the proposed CNS SAR model, specifically address SAR for secondary pharmacology with a focus on drug-induced CNS liability, presenting specific substructures associated with altered risk at the ion channels screened, using gold-standard electrophysiology data. One concern we had for this ambitious project was that there would be no detectable SAR for seizure-related ion channels. Specifically, we anticipated that the dataset could be “flat,” in that any variation of pharmacophore features away from the parent compounds would reduce potency to >100 μM, yielding little useful information. However, this was clearly not the case; the data provided a complex and detailed picture of the link between structure and potency. Indeed, many of the significant pharmacophore features were associated with increased potency; this was perhaps unexpected but advantageous in generating SAR. One caveat of this work was the relatively small compound set size which prevented application of robust machine learning approaches, such as neural networks, Random Forest models, or support vector machines. However, as an initial proof-of-concept study, this compound set was sufficient to begin developing SAR for CNS-related ion channels. To train a reliable and robust predictive SAR model would likely require ∼800 to 1000 small molecules with different chemical substructures spanning a range of IC50 values at the chosen ion channels. This is commensurate with previously published research showing datasets of ∼1000 compounds are required to train and test reliable predictive QSAR models (Roy et al. 2008; Rácz et al. 2021; National Center for Toxicological Research (NCTR) 2023; Seal et al. 2024). Further, additional ion channels could be screened to build a more complete picture across the iLASeizure panel, with a focus on the most structurally diverse channels. It would also be important to investigate which ion channels are most associated with seizure risk in the whole-cell hiPSC-neuronal model by testing a subset (∼10% of compounds) in MEA. In conclusion, the data suggest that there are key pharmacophore features that present a significant risk for potency at ion channels associated with drug-induced toxicities to the CNS. Given the relatively small dataset, care has been taken not to overinterpret the data. However, small initial clusters based on chemical structure show some commonality in terms of activity (IC50) for 2 of the 4 ion channels tested. This suggests there may be regions of chemical spaces which are privileged in terms of reducing seizure potential. Taken together, the data provide insight and a starting point for an SAR predictive tool (Fig. 7). These findings are the first report of seizure-related ion channel SAR and are the first step in helping medicinal chemists redesign their molecules to avoid CNS liability supporting optimal drug design. SAR modelling has proven successful in the cardiac liability field, with significant reductions in hERG liability and no market withdrawals due to TdP under ICH guidance (Gund 1977; Gintant et al. 2016; Garrido et al. 2020; Creanza et al. 2021). By analogy, early stage safety screening combined with an understanding of the structural features associated with ion channel potency provides an exciting opportunity to reduce CNS liability. Positioning iLASeizure SAR within the early stages of drug discovery and development. iLASeizure is currently used particularly in the lead generation and optimization phases of preclinical safety testing. An SAR tool could help evaluate and redesign compounds with a positive result in iLASeizure. 1. Compounds that have a negative result in the initial iLASeizure screen will proceed to the next stage of the discovery and development process. 2. For compounds with a positive result in iLASeizure, we propose these candidates will be applied to the iLASeizure SAR model and redesigned to remove any chemical structure flags. 3. The newly designed compound would then be screened in the iLASeizure panel, with a negative result allowing the compound to proceed to the next steps of development. 4. If the redesigned compound produced another positive result, it could be fed back into the iLASeizure SAR tool for further redesign. Supplementary Material Contributor Information Louisa A K Zolkiewski, ApconiX, Macclesfield SK10 4TG, United Kingdom. Kimberly L Rockley, ApconiX, Macclesfield SK10 4TG, United Kingdom. Ruth A Roberts, ApconiX, Macclesfield SK10 4TG, United Kingdom; Department of Biosciences, University of Birmingham, Edgbaston B15 1TT, United Kingdom. Hannah G Jennings, ApconiX, Macclesfield SK10 4TG, United Kingdom. Magali-Anne Maizières, ApconiX, Macclesfield SK10 4TG, United Kingdom. Karen Jones, ApconiX, Macclesfield SK10 4TG, United Kingdom. James L McDonagh, ApconiX, Macclesfield SK10 4TG, United Kingdom. Katie Newman, ApconiX, Macclesfield SK10 4TG, United Kingdom. Ed J Griffen, MedChemica Consultancy Ltd, Motorworks, Macclesfield, Cheshire SK11 6DU, United Kingdom. Jessica E Stacey, MedChemica Consultancy Ltd, Motorworks, Macclesfield, Cheshire SK11 6DU, United Kingdom. Michael J Morton, ApconiX, Macclesfield SK10 4TG, United Kingdom. Supplementary material Supplementary material is available at Toxicological Sciences online. Funding All work described was funded by ApconiX either directly in our own research laboratory or by paid fee for service to MedChemica. Conflicts of interest. RAR, KLR, HGJ, KJ, M-AM, LAKZ, JLMcD, KN, and M-AM are all employees of ApconiX, an integrated ion channel research and toxicology consultancy that provides expertise to academia, industry, NGOs, and government on all aspects of nonclinical program design and delivery. EJG and JES are employees of MedChemica Consultancy Ltd. Data availability All data generated or analyzed in this study are included in this published article. IC50 data for the seizure analysis (https://doi.org/10.5281/zenodo.20658862) and the ChEMBL hERG datasets (https://doi.org/10.5281/zenodo.20660831) have been deposited in Zenodo. References Associated Data Supplementary Materials Data Availability Statement All data generated or analyzed in this study are included in this published article. IC50 data for the seizure analysis (https://doi.org/10.5281/zenodo.20658862) and the ChEMBL hERG datasets (https://doi.org/10.5281/zenodo.20660831) have been deposited in Zenodo.