Uncovering molecular determinants of potency and binding affinity in hit compounds targeting FGF14/Nav1.6 complex

Abstract

Identifying molecular mechanisms that regulate neuronal excitability is essential for developing targeted therapies for neuropsychiatric disorders. The protein–protein interaction (PPI) between fibroblast growth factor 14 (FGF14) and the voltage-gated Na⁺ channel Nav1.6 is critical in regulating neuronal excitability and has emerged as a promising drug target. However, the physicochemical features that drive small-molecule modulation of this interface remain elusive. Here, we apply a descriptor-based chemoinformatics approach to analyze 15 hit compounds identified via high-throughput screening, aiming to elucidate structure–activity relationships influencing their potency and binding affinity. The analysis revealed distinct subsets of physicochemical features strongly associated with either potency or binding affinity values, suggesting that these parameters are governed by largely independent molecular determinants. This independence implies that optimizing a compound for improved affinity need not compromise potency, and vice versa. Together, these findings may guide the rational optimization of first-in-class compounds aimed at controlling neuronal excitability through targeted PPI interface modulation.

Graphical Abstract

Supplementary Information

The online version contains supplementary material available at 10.1186/s13321-025-01122-0.

Scientific contribution

This study establishes a quantitative framework for linking the chemical structure of hit compounds to modulation of the FGF14/Nav1.6 complex, a key player in neuronal excitability. By identifying specific, interpretable molecular features associated with compound bioactivity, our analysis enables informed prioritization of hit compounds for further development. The integration of chemoinformatics and statistical analysis provides a reproducible platform for rational compound optimization.

Supplementary Information

The online version contains supplementary material available at 10.1186/s13321-025-01122-0.

Introduction

The intracellular fibroblast growth factor (iFGF) family comprises FGF11–14, which function as accessory proteins of voltage-gated sodium (Nav) channels. These iFGFs interact specifically with the intracellular C-terminal domain (CTD) of Nav channels through structurally and functionally distinct protein–protein interactions (PPIs) that are unique to each iFGF and Nav channel isoform [1]. Unlike conventional PPI interfaces, which are typically shallow and lack well-defined binding sites [2], the iFGF–Nav CTD interfaces feature distinct, druggable pockets that are amenable to pharmacological targeting [3–8]. Among the iFGF-Nav CTD complexes, the interaction between FGF14 and Nav1.6 has gained attention as a potential druggable target for fine-tuning Na⁺ currents and neuronal excitability in the nucleus accumbens (NAc) [1–5], a region central to reward processing and implicated in various neuropsychiatric disorders [6, 7]. By modulating neuronal firing patterns of medium spiny neurons (MSN) in the NAc with a mechanism mediated by PPI, targeting the FGF14/Nav1.6 interface may offer a novel precision medicine strategy for addressing circuit-level dysfunctions associated with conditions such as major depressive disorder, substance use disorder, and other psychiatric illnesses [1, 8] that currently lack targeted treatments.

Through a combination of high-throughput screening (HTS) and orthogonal assay cascades, small-molecule modulators were recently identified and characterized for their ability to modulate FGF14/Nav1.6 interaction [9]. An initial bioluminescence-based screen of approximately 45,000 small molecules identified 16 hit compounds for further study. These compounds demonstrated reproducible potency in the split-luciferase complementation assay (LCA) and measurable binding affinity to FGF14 and/or the Nav1.6 CTD in surface plasmon resonance (SPR) assays. One hit compound was advanced for hit-to-lead optimization and successfully refined for in vivo efficacy studies, while the remaining hits have yet to be explored [9]. These scaffolds offer a valuable foundation for hit expansion and further lead optimization toward drug candidate selection. Meaningful progress in this direction requires structure–activity relationship (SAR) insights that extend beyond the PPI interface, encompassing broader molecular features that influence the potency and selectivity of these small-molecule modulators [10].

Lead optimization efforts often rely on molecular docking and molecular dynamics simulations to visualize potential binding poses. However, these approaches are frequently limited by a lack of mechanistic insight into the structural or electronic properties that drive strong protein binding and functional modulation [11, 12]. Chemoinformatic analysis offers a complementary, data-driven strategy to identify key chemical features of small molecules correlated with binding affinity and functional outcomes, enhancing the efficiency of lead optimization beyond inspection of the druggable interface [13].

To address this critical gap, chemoinformatics approaches grounded in quantitative structure–activity relationship (QSAR) modeling have emerged as powerful tools for establishing a relationship between the molecular structure of small molecules and their bioactivity [14–18]. By converting chemical structures into interpretable numerical descriptors and applying statistical or machine learning methods, QSAR enables the identification of physicochemical features predictive of biological function [16, 19–22]. Such chemoinformatics strategies offer scalable, data-centric pipelines capable of revealing nonlinear patterns and hidden determinants of functional potency and binding affinity across diverse chemical scaffolds [16]. To fully realize the potential of rational drug design, these computational approaches have become indispensable for accelerating hit selection, hit expansion and lead optimization, improving the predictive power of molecular screening [23].

On the chemoinformatics front, significant progress has been made with the development of modern molecular descriptors that transcend simple scalar features such as atom counts, molecular weight, or bond numbers rather capturing more sophisticated molecular characteristics, such as polarizability, electronic distribution, and topological complexity, that play a decisive role in modulating both binding affinity and functional inhibition [14, 21]. A number of descriptor generation tools have been developed to capture such complexity, providing an extensive set of features designed to characterize molecular structure across multiple dimensions [24–28]. In the context of PPI modulators, leveraging a broad and diverse set of chemical descriptors becomes essential for capturing the nuanced features that drive binding affinity and functional potency.

Here, we performed a chemoinformatic analysis of validated hits from the FGF14/Nav1.6 HTS campaign, with the goal of identifying statistically significant correlations between molecular features and experimental potency and binding affinity values. We calculated a comprehensive set of molecular descriptors using the Mordred package based on SMILES representations of each molecule. These descriptors spanned constitutional, topological, geometric and electronic properties and physicochemical features [21]. Each descriptor was systematically analyzed for its statistical correlation with experimentally measured IC₅₀ (potency) and K_d (binding affinity) values for FGF14/Nav1.6 modulation. Features showing strong and statistically significant Spearman’s rank correlation coefficients were prioritized, enabling the identification of interpretable molecular determinants of FGF14/Nav1.6 modulation.

Materials and methods

Dataset

We utilized a curated dataset comprising 16 small-molecule hits identified through high-throughput screening (HTS) targeting the FGF14/Nav1.6 PPI interface, as described in Ref. [9]. The initial compound pool (~ 45,000 molecules) was sourced from the Maybridge and ChemBridge chemical libraries and screened using a split-luciferase complementation assay (LCA) designed to detect modulators of the FGF14/Nav1.6 complex in living cells. Following the primary screen, compounds were filtered through cytotoxicity assessment, counter-screening against full-length luciferase to eliminate reporter-interfering hits, structural filtration to exclude PAINS (Pan-Assay Interference Compounds) using SwissADME [29], and manual curation based on drug-likeness criteria (e.g., Lipinski’s Rule of Five). All selected compounds were re-solubilized from fresh powder, and purity was confirmed at > 98% by HPLC analysis. This process yielded a refined pool of 70 compounds, which were subsequently validated through LCA-based dose–response assays. Of these, 16 compounds demonstrated reproducible, dose-dependent inhibition of FGF14/Nav1.6 complex assembly, with half-maximal inhibitory concentrations (IC₅₀) ranging from 0.97 µM to 21.78 µM and binding affinity (K_d) ranging from 0.94 µM to 220 µM. These compounds were further validated in orthogonal assays, including protein thermal shift (PTS) to assess changes in protein stability and surface plasmon resonance (SPR) to quantify direct binding to FGF14 and the Nav1.6 C-terminal domain (CTD). The resulting set of validated compounds served as the basis for our chemoinformatic analysis of molecular determinants of binding affinity and potency (see Table 1 for compound details and structures). Visualizations of all 16 hit compounds and their SMILES representations are provided in Fig. S1 and Table S1, respectively. Furthermore, SPR results for 5 of the 16 compounds were reported previously (Ref. 9); for the remaining 11 compounds, the corresponding SPR sensorgrams, steady-state isotherms, and curve fits are provided in the Supplementary Information (Figures S2–S3).

Table 1.

Potency and binding affinity values for 16 hit compounds. Data adapted from Ref. [9]. ND = Not Determined, meaning the compound was tested but a reliable fit could not be obtained

Compound ID	IC50 for FGF14/Nav1.6 (μM)	${\text{K}}_{\text{d}}$ for FGF14 (μM)	${\text{K}}_{\text{d}}$ for Nav1.6 (μM)
5210801	9.33	ND	9.163
5335477	6.32	212	72.84
5671129	21.78	150	ND
5674122	0.97	0.94	4.07
5936750	5.64	13.57	ND
6039152	11.69	220.7	17.22
6465241	7.01	ND	7.73
6475665	12.79	104.7	122.4
7605086	0.99	1.9	9.83
7630101	12.19	114.9	7.277
7634112	15.81	66.47	ND
7647895	10.69	8.53	3.96
7785745	12.85	24.99	ND
BTB11798	3.13	ND	ND
CC18209	2.16	1.634	ND
SPB02176	10.52	ND	ND

Extraction of molecular descriptors

Molecular descriptors quantitatively extract structural, physicochemical, and electronic properties of chemical compounds, enabling their transformation into high-dimensional numerical vectors suitable for computational analysis. Descriptors can be categorized by the dimensionality of the structural data they capture [23]. Zero- and one-dimensional descriptors represent simple scalar features, often derived from experimental measurements or basic molecular properties, such as molecular weight, atom counts, and counts of functional groups [21]. Two-dimensional descriptors encode molecular topology through graph-based features such as connectivity indices, atom-pair correlations, and fingerprint-like patterns that remain invariant under atom renumbering. These are widely used in cheminformatics due to their computational efficiency and ability to capture relevant structural information directly from SMILES input. In contrast, three-dimensional descriptors reflect spatial and geometric properties, such as molecular volume, surface area, dipole moment, and conformational distributions, which often require computed or experimentally derived 3D structures.

In this work, descriptor calculation was conducted using the Mordred package version 1.2.0, which can compute over 1800 features across multiple descriptor classes [28]. However, due to the absence of 3D conformers and occasional calculation failures or unsupported features, only 1428 descriptors were successfully generated and retained for analysis. These included constitutional descriptors, 2D topological indices, and a broad array of graph-based features, including a large subset of 606 autocorrelation descriptors.

Cosine similarity

Chemical similarity, otherwise known as molecular similarity, has long been widely applied in ligand-based drug discovery, operating under the intuitive assumption that structurally similar molecules will exhibit similar biological activities [30–33]. Traditional approaches such as the Tanimoto coefficient, often applied to 2D molecular fingerprints, offer efficient ways to estimate structural overlap but typically lack resolution in capturing subtle physicochemical or topological distinctions [34]. This highlights the need for interpretable approaches that can detect descriptor-level variations among structurally related compounds. Among several similarity measures available, cosine similarity is known for its effectiveness in comparing directional alignment of feature vectors [34]. However, because the raw descriptors have different numerical scales, normalization is a critical preprocessing step to ensure comparability. Thus, we rescaled all feature values to a common range between 0 and 1 using min–max normalization, as defined by the equation:

$$\widehat{z}=\frac{z-z_{\mathit{min}}}{z_{\mathit{max}}-z_{\mathit{min}}}$$ 1

where $\widehat{z}$ is the normalized feature value, z is the original (unnormalized) feature value, and z_min and z_max are the minimum and maximum of that feature across all compounds, respectively. To compare the physicochemical profiles of two molecules, A and B, we represent them as vectors in an n-dimensional feature space: A = [A₁, A₂, …, Aₙ] and B = [B₁, B₂, …, Bₙ]. The cosine similarity between them is then calculated using the formula:

$$\text{Similarity}\left(\text{A},,,\text{B}\right)=\text{cos}(\mathrm{\Theta })=\frac{{\sum }_{i=1}^n{A}_i{B}_i}{\sqrt{\sum _{i=1}^n{A}_i^2}\sqrt{\sum _{i=1}^n{B}_i^2}}$$ 2

where Θ is the angle between vectors A and B in the feature space. We note that after normalizing all feature values to a common scale between 0 and 1, the resulting cosine similarity is constrained within the range [0, 1]. A value of 1 indicates complete alignment between vectors, likely indicating similar molecular features. A value of 0 indicates orthogonality, i.e., no similarity between the feature profiles.

Feature selection

While global similarity measures such as cosine similarity offer a comprehensive comparison across high-dimensional descriptor space, they often fail to disentangle which specific molecular features are responsible for bioactivity. To identify which molecular descriptors most strongly influence functional potency or binding affinity, we computed Spearman rank correlation coefficients between each of the Mordred descriptors and the experimentally measured IC₅₀ and K_d values for FGF14/Nav1.6 inhibition. This approach has been shown to identify monotonic relationships between individual molecular descriptors and the experimentally measured bioactivity [35]. This method not only allows for an interpretable SAR, but also addresses a central limitation of whole-profile similarity: two molecules may appear highly similar in descriptor space yet exhibit divergent bioactivity due to one or two distinct molecular features.

The Spearman’s rank correlation coefficient, defined as the Pearson correlation between the ranks of two variables, is given by:

$$r_s=\frac{⟨R_X{R}_Y⟩-⟨R_X⟩⟨R_Y⟩}{\sqrt{⟨R_X^2⟩-{⟨R_X⟩}^2}\sqrt{⟨R_Y^2⟩-{⟨R_Y⟩}^2}}$$ 3

where R_X and R_Y are the ranks of the molecular descriptor and bioactivity (IC₅₀ or K_d), respectively. It is important to note that because the Spearman coefficient depends on the variance of the ranks, it is undefined when all values for a descriptor are identical or if the ranking has no variation. This means that the Spearman coefficient is undefined for those descriptors that are constant across all compounds or show no meaningful spread in values.

Results

Due to the presence of two discrete modules in compound 6039152, Mordred package was unable to compute molecular descriptors for this molecule. Therefore, our chemoinformatic analysis was limited to the remaining 15 validated compounds. Among the 15 validated hit compounds identified through HTS, we observed pronounced divergence in both potency and binding selectivity across the FGF14/Nav1.6 interface. Based on SPR and PTS assays, several compounds demonstrated selective binding to FGF14, while others exhibited exclusive binding to the Nav1.6 CTD or dual-target engagement with both proteins [9]. Notably, this diversity emerged despite substantial scaffold similarity among the compounds, highlighting the influence of subtle physicochemical and topological differences in shaping bioactivity. These findings underscore the importance of detailed molecular profiling for understanding target engagement mechanisms and guiding structure-based optimization.

To visualize the overall chemical descriptor diversity across 15 hit compounds, we generated a heatmap of normalized molecular features computed via the Mordred package (Fig. 1). Each row represents a compound, and each column corresponds to one of the 1428 numerical descriptors. For clarity, descriptors were grouped into two categories i.e. autocorrelation functions (606) and other features (822). The left portion of the heatmap displays autocorrelation descriptors, which reflect topological and spatial relationships among atoms, while the right portion contains other 2D molecular descriptors representing a variety of constitutional, physicochemical, and fragment-based features. Darker blue regions correspond to higher descriptor values, while lighter regions indicate lower normalized intensities. This visualization reveals that most descriptors are shared across compounds, with only a limited subset of autocorrelations, predominantly on the left side of the heatmap, exhibiting significant variability. Moreover, compound CC18209 displays consistently low descriptor intensity, highlighting its minimalist chemical profile across nearly all feature classes. In contrast, compounds such as 7647895 and 5674122 exhibit uniformly high intensities across numerous descriptors, suggesting dense and chemically rich scaffolds. Compound 7605086, which was previously selected for an optimization campaign [9], presents a balanced and moderately complex descriptor profile. Overall, this heatmap provides a global overview of feature heterogeneity within the dataset.

Fig. 1.

Heatmap of normalized Mordred descriptors across 15 hit compounds, grouped by autocorrelations (606 descriptors) and other 2D molecular descriptors (822 descriptors)

Next, we utilized cosine similarity over descriptor space to provide a scalable and information-rich representation of molecular similarity. The results for cosine similarity of compounds are presented in Fig. 2a. Notably, compound CC18209 exhibits the lowest average similarity to the rest of the compound set, highlighting its structural divergence among the selected hits. This deviation likely stems from its compact size, limited functional group diversity, and reduced topological complexity, an observation that aligns with the sparse descriptor profile seen in Fig. 1. In contrast, compounds such as 6465241 and 7785745 exhibit consistently high similarity to the rest of the dataset, suggesting the presence of conserved scaffolds or common descriptor patterns across multiple compounds. To quantify this more precisely, we calculated the average cosine similarity for each compound with all other compounds in descriptor space (Fig. 2b). The resulting values range from 0.53 to 0.81.

Fig. 2.

a Heatmap of pairwise cosine similarity between compounds. Darker blue values indicate higher similarity, and lighter yellow indicates lower similarity. b Barplot of average cosine similarity for each compound. Compounds are ordered from highest to lowest average similarity

Having characterized the similarity between compounds, we next aimed to identify which molecular features are correlated with their potency and binding affinity. The distribution of correlation coefficients for IC₅₀ is shown in Fig. 3a, revealing a wide spread of descriptor–activity relationships. To focus on the most informative molecular features, we applied a threshold of |r_s|≥ 0.7, identifying a shortlist of descriptors with statistically significant correlations (p < 0.005). These are summarized in Fig. 3b, which highlights both positively and negatively correlated features. The definition of these features is provided in Table 2. Notably, descriptors such as MATS1i, AATSC1i, and ATSC1i, BCUTdv-1l exhibit strong negative correlations with IC₅₀, while SssCH2, SMR_VSA6, and SlogP_VSA2 show strong positive correlations. Similar descriptors were previously identified in the context of selective Nav1.2 and Nav1.6 inhibition, relevant to potential treatments for Dravet Syndrome [36].

Fig. 3.

a Distribution of Spearman correlation coefficients across all molecular descriptors in relation to IC₅₀ values for FGF14/Nav1.6. b Ranked importance of molecular descriptors exhibiting strong correlations with IC₅₀ values, selected using a threshold of |r_s|≥ 0.7. c The averaged and centered Moreau–Broto autocorrelation at lag 1, weighted by the first ionization potential (ATSC1i), exhibits a significant negative Spearman correlation with IC₅₀ (r_s = − 0.77, p = 0.00076). d SMR_VSA6, a MOE-type van der Waals surface area descriptor representing the solvent-accessible area of atoms with molar refractivity between 2.75 and 3.05, shows a positive Spearman correlation with IC₅₀ (r_s = 0.75, p = 0.00127)

Table 2.

Definition of selected features that strongly correlated with IC₅₀ values

Molecular descriptor	Definition
ATSC1i	Centered moreau-broto autocorrelation of lag 1 weighted by ionization potential
AATSC1i	Averaged and centered moreau-broto autocorrelation of lag 1 weighted by ionization potential
MATS1i	Moran coefficient of lag 1 weighted by ionization potential
BCUTdv-1 l	First highest eigenvalue of Burden matrix weighted by valence electrons
SssCH2	Sum of ssCH2
SMR_VSA6	MOE MR VSA Descriptor 6 (2.75 <  = x < 3.05)
SlogP_VSA2	MOE logP VSA Descriptor 2 (− 0.40 <  = x < − 0.20)

To further analyze these relationships, Fig. 3c and Fig. 3d show scatter plots for ATSC1i and SMR_VSA6, respectively, against IC₅₀ values. Both descriptors exhibit clear monotonic trends and absolute Spearman correlation coefficients around 0.75. ATSC1i, defined as the centered Moreau–Broto autocorrelation of lag 1 weighted by ionization potential, captures the distribution of ionization potential across atom pairs separated by one bond in the molecular graph. In this context, positive ATSC1i values indicate that atoms one bond apart tends to have similar ionization potentials, reflecting local electronic homogeneity, while negative values indicating localized electronic heterogeneity. Compounds with positive ATSC1i tend to have high potency (low IC₅₀), while those with lower or negative values show weaker activity. This suggests that a uniform distribution of ionization potential across atoms one bond apart may enhance molecular recognition by reducing local electronic mismatch, thereby promoting stable electrostatic or dipolar interactions at the FGF14/Nav1.6 binding interface.

Another selected feature, SMR_VSA6, belongs to the MOE-type van der Waals surface area (VSA) family, which partitions a molecule’s surface area according to predefined ranges of molar refractivity [37]. Specifically, SMR_VSA6 quantifies the total solvent-accessible surface area contributed by atoms whose molar refractivity values fall between 2.75 and 3.05, a range often associated with moderately polarizable, weakly electron-rich atoms. The positive correlation observed between SMR_VSA6 and IC₅₀ (r_s = 0.75) suggests that compounds with greater surface area in this refractivity range tend to exhibit lower potency. This may imply that an overrepresentation of moderately polarizable surface regions could interfere with optimal molecular recognition or binding efficiency at the FGF14/Nav1.6 interface, potentially by reducing electrostatic complementarity or introducing unfavorable steric effects.

SlogP_VSA2 is a descriptor within the MOE logP-based van der Waals surface area (VSA) family, designed to quantify a molecule’s surface area contribution from atoms according to their predicted atomic logP values. logP is a measure of a compound’s lipophilicity, indicating its relative solubility for lipid versus aqueous environments. A positive LogP value (LogP > 0) suggests the compound is lipophilic (lipid-soluble), while a negative LogP value (LogP < 0) indicates the compound is hydrophilic (water-soluble). Specifically, SlogP_VSA2 measures the total solvent-accessible surface area of atoms with logP contributions between –0.40 and –0.20, representing weakly hydrophilic character. The observed positive correlation with IC₅₀ values (Spearman’s r_s = 0.74) indicates that compounds with larger SlogP_VSA2 values tend to exhibit lower potency. This suggests that an excess of weakly polar surface exposure may interfere with favorable hydrophobic or electrostatic interactions at the FGF14/Nav1.6 interface, potentially disrupting optimal binding or molecular recognition.

Additionally, BCUTdv-1 l, which represents the largest eigenvalue of the Burden matrix weighted by valence electrons, shows a strong negative correlation with IC₅₀ values (Spearman’s r_s ≈ –0.7). High BCUTdv-1 l values suggest the presence of extended, electron-rich regions and enhanced conjugation or polarizability, which may contribute to stronger binding affinity. Notably, the relationship appears to exhibit a threshold-like behavior, where compounds with BCUTdv-1 l values above ~ 0.6 consistently display low IC₅₀ values, while those below this threshold tend to show weaker potency. This suggests that a minimum level of valence-electron-weighted connectivity may be required to enable effective interactions at the binding site, potentially via long-range electronic effects or improved charge delocalization.

Another pivotal parameter for understanding drug-target interaction is binding affinity or dissociation constant (K_d) of hit compounds versus FGF14. This parameter quantifies how tightly a compound associates with its molecular target under equilibrium conditions. A lower K_d reflects stronger binding, signifying that a compound can occupy the target site even at low concentrations. For weak binders (> 100 µM, i.e., high K_d), the curves were adequately fit, with sufficient data points approaching a plateau. All fits followed a sigmoidal shape using a 1:1 binding model, and the signal-to-noise ratio was high near the plateau. The 95% confidence intervals on the K_d were based on several (n = 3–5) repeated experiments. As shown in Fig. 4a, Spearman rank correlation analysis revealed a wide range of descriptor–affinity relationships. Applying a threshold of |r_s|≥ 0.7. and p < 0.005 identified several descriptors with statistically significant associations, summarized in Fig. 4b. Notably, all of these features are autocorrelation-based descriptors, defined in detail in Table 3.

Fig. 4.

a Distribution of Spearman correlation coefficients across all molecular descriptors in relation to K_d values for FGF14. B Ranked importance of molecular descriptors exhibiting strong correlations with K_d values, filtered using a Spearman threshold of |r_s|≥ 0.7. c Centered Moreau–Broto autocorrelation at lag 4, weighted by atomic number (designated as AATSC4Z in Mordred), exhibits a strong negative Spearman correlation with K_d (r_s = − 0.78, p = 0.00447). d VSA_EState3, a descriptor combining van der Waals surface area with atoms whose electrotopological state falls within the range 5.00 ≤ x < 5.41, shows a strong negative Spearman correlation with K_d (r_s = − 0.81, p = 0.00256)

Table 3.

Definition of selected features that strongly correlated with K_d for FGF14

Molecular descriptor	Definition
AATSC4Z	Averaged and centered moreau-broto autocorrelation of lag 4 weighted by atomic number
AATSC4m	Averaged and centered moreau-broto autocorrelation of lag 4 weighted by mass
MATS4c	Moran coefficient of lag 4 weighted by gasteiger charge
C1SP2	SP2 carbon bound to 1 other carbon
VSA_EState3	VSA EState Descriptor 3 ( 5.00 <  = x < 5.41)
nHRing	Hetero ring count

To elucidate molecular features that influence binding affinity, we examined two top-ranked descriptors identified through Spearman correlation analysis (Fig. 4c, d). AATSC4Z is the averaged and centered Moreau–Broto autocorrelation of lag 4, weighted by atomic number (Z). This descriptor quantifies how atomic numbers are distributed across pairs of atoms that are four bonds apart. The atomic number defines an element’s identity and its nuclear charge, influencing many atomic properties. In general, atoms with higher atomic numbers are heavier, more polarizable, and exhibit greater electron-attracting capacity compared to lighter elements with lower atomic numbers. Negative AATSC4Z values indicate an inhomogeneous distribution, in particular, frequent pairing of heavy and light atoms (e.g., Sulfur and carbon, or iodine and nitrogen) at that distance. Notably, compound 6475 has the lowest AATSC4Z value (–6), which can be attributed to the presence of iodine, making it the only hit compound with such a heavy atom. Such structural inhomogeneity may introduce asymmetries in electronic distribution and conformational flexibility, which can reduce electrostatic and steric complementarity with the protein interface. This is consistent with the observed negative correlation between AATSC4Z and binding affinity (r_s = –0.78, p = 0.00447) as shown in Fig. 4c.

VSA_EState3, on the other hand, quantifies the total van der Waals surface area of atoms in a molecule whose electrotopological state (EState) values fall within a defined window (5.00 ≤ x < 5.41). This descriptor effectively captures the contribution of moderately electron-rich atoms, such as oxygen or nitrogen, to the molecule’s exposed surface area. A strong negative correlation (r_s = –0.81, p = 0.00256) indicates that exposed, moderately electron-rich atoms within this EState range are strongly associated with enhanced binding of a compound to FGF14 (Fig. 4d). Such atoms may engage in polar interactions (e.g., dipole–dipole or hydrogen bonding) or promote desolvation at the PPI interface, thereby lowering K_d and stabilizing the complex. Notably, compound 7630101 has a VSA_EState3 value of zero, indicating a complete absence of exposed, moderately electron-rich atoms (e.g., oxygen or nitrogen) within the defined EState range (5.00 ≤ x < 5.41). The absence of these atoms likely eliminates functional groups capable of strong polar interactions or favorable desolvation, consistent with the compound’s weaker binding affinity. Interestingly, the saturation-like trend indicates that once a critical number of exposed and moderately electron-rich atoms is reached, additional contributions no longer improve binding of a compound with FGF14 interface. This plateau likely arises because the available binding interface becomes fully engaged, limiting further electrostatic or desolvation gains despite an increased presence of favorable atoms.

Next, we analyzed the molecular features associated with compound binding affinity to the Nav1.6 CTD. The results are presented in Fig. 5. Unlike our analysis of FGF14 binding, the features associated with Nav1.6 K_d values are more diverse, belonging to different Mordred feature categories, with all descriptors surpassing the Spearman correlation threshold (|r_s|≥ 0.8). Broadly, the most predictive features fall into three conceptual domains: Majority of the features that strongly correlate with Nav1.6 K_d are autocorrelations, such as ATS(n)se, ATS(n)pe, ATS(n)are, and ATS(n)i, which reflect how atomic properties like electronegativity, and ionization potential are distributed across the molecular structure at various topological distances. Molecular descriptors such as SpMax_D, SpMAD_D, LogEE_D, and SpDiam_Dzm, encode higher-order molecular shape, symmetry, and flexibility by analyzing eigenvalues of matrices derived from atomic distances or Barysz transformations. While constitutional features such as nBondsKS, nAtom, nBonds, and Kier1, capture molecular size and branching complexity. The definition of the selected features is provided in Table 4.

Fig. 5.

a) Distribution of Spearman correlation coefficients across all molecular descriptors in relation to K_d values for Nav1.6. B Ranked importance of molecular descriptors exhibiting strong correlations with K_d values, filtered using a Spearman threshold of |r_s|≥ 0.7. c Broto–Moreau autocorrelation at lag 5, weighted by ionization potential (ATS5i), exhibits a strong positive Spearman correlation wit K_d (r_s = − 0.9, p = 0.00201). d Total number of bonds (non-kekulized structure), nBonds, shows a strong positive Spearman correlation with K_d (r_s = − 0.9, p = 0.00244)

Table 4.

Definition of selected features that strongly correlated with K_d for Nav1.6

Molecular descriptor	Definition
ATS(n)se	Moreau–Broto autocorrelation of lag n, weighted by Sanderson electronegativity
ATS(n)pe	Averaged and centered Moreau–Broto autocorrelation of lag n, weighted by Pauling electronegativity
ATS(n)are	Moreau–Broto autocorrelation of lag n, weighted by Allred–Rochow electronegativity
ATS(n)i	Moreau–Broto autocorrelation of lag n, weighted by ionization potential
SpDiam_D	Spectral diameter from the topological distance matrix
SpMax_D	Leading eigenvalue (spectral maximum) of the topological distance matrix
SpMAD_D	Spectral mean absolute deviation from the distance matrix
SpAD_D	Spectral absolute deviation from the distance matrix
SpAbs_D	Graph energy (absolute eigenvalue sum) from the distance matrix
LogEE_D	Estrada-like topological index from the distance matrix
SpDiam_DzZ	Spectral diameter from Barysz matrix, weighted by atomic number
SpMAD_DzZ	Spectral mean absolute deviation from Barysz matrix, weighted by atomic number
SpDiam_Dzm	Spectral diameter from Barysz matrix, weighted by atomic mass
SpMAD_Dzm	Spectral mean absolute deviation from Barysz matrix, weighted by atomic mass
SpDiam_Dzi	Spectral diameter from Barysz matrix, weighted by ionization potential
Sse / Spe / Sare / Si	Sum of Sanderson, Pauling, Allred–Rochow electronegativities, and ionization potentials, respectively, across all atoms
nAtom	Total number of atoms in the molecule
nBonds	Total number of bonds (non-kekulized structure)
nBondsKS	Number of single bonds in kekulized structure
Kier1	Kappa shape index 1, a measure of molecular cyclicity and flexibility

We focus on two molecular descriptors with strong correlations to binding affinity with Nav1.6. The first, ATS5i, the Moreau–Broto autocorrelation of lag 5 weighted by ionization potential, exhibited a significant negative correlation with K_d (r_s = − 0.9, p = 0.00201) as shown in Fig. 5c. ATS5i quantifies how atoms with high or low ionization potential are distributed across topological distances of five bonds within the molecule. A higher ATS5i value indicates a more homogeneous and extended distribution of ionization potential across the molecular scaffold. The negative correlation of ATS5i with K_d values suggests that such a regular distribution facilitates electrostatic complementarity and charge delocalization, which may be critical for engaging the charged or polar regions of the Nav1.6 CTD binding interface.

The second descriptor, nBonds, represents the total number of chemical bonds in the molecule based on its non-kekulized structure, that is, without assigning fixed aromatic resonance patterns. In this representation, all bonds are treated uniformly (e.g., aromatic bonds are not averaged or delocalized), allowing for a more consistent assessment of overall molecular complexity. Thus, a higher nBonds value reflects increased molecular size, topological diversity, and a greater number of potential contact points or interaction sites with the protein surface. This descriptor exhibited a strong negative correlation with K_d (r_s = –0.9, p = 0.00244), as shown in Fig. 5d. Molecules with fewer bonds tended to exhibit weaker binding affinity (higher K_d) while those with a greater number of bonds showed markedly stronger binding. This relationship suggests that compounds must exceed a certain threshold of structural complexity or molecular size to engage the Nav1.6 CTD interface effectively.

To identify which molecular descriptors most effectively capture the pharmacological differences between these compounds, we calculated the pairwise correlation among descriptors predictive of K_d for FGF14 and IC₅₀ for FGF14/Nav1.6, ensuring that redundant or highly collinear features did not disproportionately bias interpretation (Fig. 6). Importantly, descriptors predictive of binding affinity and potency displayed very weak cross-correlation, suggesting that these parameters are governed by largely independent molecular determinants. This independence signifies that optimizing a compound for enhanced binding might not necessarily compromise its potency, and vice versa.

Fig. 6.

Pearson correlation matrix of molecular descriptors predictive of K_d for FGF14 and IC₅₀ values for FGF14/Nav1.6. Grid colors indicate the strength of correlation (r), while p-values are annotated in each grid to denote statistical significance. IC₅₀-related features and K_d -related features are labeled in blue and green, respectively

In the original study by Dvorak et al., compound 7605086-referred to as 7605 throughout the rest of the study-was identified as a potent dual binder to FGF14 and Nav1.6, displaying favorable nanomolar affinity [9]. However, despite its promising pharmacological profile, 7605 suffered from poor aqueous solubility, limiting its in vivo potential. To overcome this limitation, a rational modification strategy was employed: the lipophilic 3-methyl group on the phenyl ring of 7605 was replaced with a 3-carboxylic acid ethyl ester moiety, yielding the analogue 1028 (aka PW1028, Fig. 7a, b). This transformation preserved key binding motifs while improving drug-like properties, including solubility and blood–brain barrier (BBB) permeability. Notably, the substitution also resulted in altered target engagement: 1028 retained high-affinity binding (K_d = 0.293 µM) to FGF14 but lost detectable interaction with Nav1.6, achieving a shift in selectivity. In addition, 1028 exhibited enhanced potency, with an IC50 of 0.247 µM for FGf14/Nav1.6.

Fig. 7.

a Chemical structure of the parent compound 7605. b Chemical structure of the optimized analogue 1028. c Radar plot comparing 7605 and 1028 across molecular descriptors most strongly correlated with potency (IC₅₀). d Radar plot comparing 7605 and 1028 across molecular descriptors most strongly correlated with binding affinity (K_d)

To better understand the molecular determinants of this functional transformation, we performed a comparative descriptor analysis of the parent compound 7605 and its optimized analogue 1028. This analysis reveals how small structural changes are reflected in descriptor values. First, we examined key descriptors previously associated with potency, focusing on their ability to differentiate the two molecules based on relevant physicochemical properties. These include one autocorrelation descriptor weighted by ionization potential (ATSC1i), one topological feature (SssCH₂), and two surface area-based descriptors (SMR_VSA6 and SlogP_VSA2). The radar plot (Fig. 7c) reveals marked differences between the two compounds. Compound 1028 exhibits elevated SMR_VSA6 and SlogP_VSA2 values, consistent with the introduction of a polarizable ester moiety. Interestingly, ATSC1i is more negative in the less potent parent compound, which deviates from the broader trend in the full dataset where negative ATSC1i generally correlates with higher IC₅₀. This discrepancy may reflect the complexity of multi-parameter optimization, where potency is shaped by a balance of features rather than a single dominant descriptor.

Figure 7d presents a radar plot comparing compound 7605086 with its analogue PW1028 across four molecular descriptors most strongly associated with binding affinity for FGF14. The values of VSA_EState3, which captures the exposed surface area of moderately electron-rich atoms, are nearly identical between the two compounds, indicating preservation of this surface feature. However, more prominent differences are observed in AATSC4Z and AATSC4m, which are Broto–Moreau autocorrelation descriptors weighted by atomic number and atomic mass, respectively. These descriptors are significantly more negative in PW1028, likely due to the presence of two additional oxygen atoms (Fig. 7b). As discussed in methods, the heavier atomic mass and localized electronegativity of oxygen increase the inhomogeneity in atomic charge and mass distribution, which in turn makes these autocorrelation values negative. Furthermore, the identical value of nHRing (number of aromatic heterocycles) in both compounds confirms that the aromatic heterocyclic scaffold is preserved, supporting the idea that the core structure remains unchanged during optimization. While the radar plot highlights notable electronic and topological differences between analogues, these do not fully follow the trends observed in feature–activity correlations across the broader dataset. Nevertheless, such analyses remain valuable for revealing how localized structural changes influence individual molecular properties, offering complementary insight for analogue design alongside statistical feature selection.

Discussion

In this study we utilized cheminformatics analysis to elucidate the SAR for hit compounds targeting the FGF14/Nav1.6 PPI interface by integrating chemical descriptors with experimental bioactivity parameters such as potency and binding affinity. Our analysis suggests that potency and binding affinity are associated with largely distinct sets of physicochemical features, indicating that different properties may differentially influence PPI inhibition versus target engagement and could inform downstream lead optimization. By examining the molecular descriptor landscape, similarity measures, and correlation analyses, we uncovered key molecular signatures associated with both potency and affinity. These findings provide a framework for interpreting bioactivity, highlighting the underlying diversity among structurally related compounds.

Our analysis indicates that a combination of various 2D molecular features including autocorrelation as well as electrotopological and surface area features exhibit a moderate-to-strong correlation with IC₅₀ values. Autocorrelation descriptors, which capture molecular connectivity patterns and electronic distributions, may contribute to inhibition of PPI complex formation by shaping the conformational accessibility of the ligand in the active site. Electrotopological features, on the other hand, may influence intermolecular electronic interactions that determine inhibitory strength. Additionally, surface area characteristics appear to play an important role in governing molecular accessibility to the binding pocket, affecting ligand-target interactions that modulate potency. Together, these features reflect a complex interplay between localized electronic structure and overall molecular footprint in governing inhibitory potency.

Similarly, binding affinity correlates most with autocorrelation descriptors in our data, suggesting that target engagement is primarily governed by atomic-level electronic distribution. For instance, AATSC4Z and AATSC4m, autocorrelation descriptors weighted by atomic number and atomic mass, respectively, exhibited strong correlations with binding affinity, indicating that heterogeneity in atomic distribution (e.g., due to heavy atoms like oxygen) contributes to weaker binding affinity. Additionally, VSA_EState3, which integrates surface area with partial charge, and nHRing (number of heteroaromatic rings) also ranked among the top features, emphasizing the roles of electronic surface complementarity and number of rings in promoting binding affinity. These findings suggest that descriptors capturing localized electronic environments and molecular architecture are important determinants of binding strength. Notably, for VSA_EState3, the correlation with binding affinity appears to plateau beyond a certain threshold. This behavior might suggest that once a sufficient level of this feature is reached, further increases do not proportionally enhance the thermodynamic stability of the ligand–target complex.

Furthermore, the distinct descriptors associated with IC₅₀ and K_d can be utilized for rational compound optimization. Although the features most strongly correlated with potency and binding affinity show minimal overlap, they point to different molecular properties that can be independently modulated. However, our comparative analysis of compound 1028 and its parent compound 7605 reveals that the values of selected descriptors most strongly correlated with potency (e.g., ATSC1i, SMR_VSA6, SlogP_VSA2, SssCH₂) do not consistently follow the expected monotonic trends observed across the full dataset. This discrepancy highlights that global descriptor–activity correlations may not always translate directly to individual lead optimization cases, especially when structural modifications introduce nonlinear or compensatory effects. Despite these limitations, the selected descriptors remain useful as tunable molecular parameters. When integrated with medicinal chemistry knowledge, they can inform iterative design strategies that balance target engagement with favorable physicochemical properties, which could support the development of analogues with improved therapeutic potential.

While our descriptor-based framework provides valuable insight into the bioactivity of the available hit compounds, it is important to highlight the limitations of our method. First, the analysis relies on 2D molecular descriptors and does not fully capture conformational flexibility or explicit 3D pharmacophoric alignments. These are often important for molecular docking studies and free-energy calculations, where spatial complementarity and dynamic adaptations play a significant role in ligand–target recognition. Additionally, the dataset is limited in size, which may constrain the statistical power and generalizability of the observed correlations. Future work should incorporate larger libraries of structural analogues, particularly around promising scaffolds such as compound 7605 and its optimized analogue 1028, to validate and refine these descriptor–bioactivity relationships. Furthermore, the integration of machine learning models trained on augmented descriptor sets, including 3D molecular features, may improve predictive accuracy, enabling a more comprehensive understanding of structure–activity trends.

Author contributions

H.T., and F.L. designed the research. J.Z., C.Z. and H. C. provided chemical support and generated new analogues such as PW1028. H.T. performed chemoinformatic analysis. A.K.S. and P.A.W analyzed results from binding studies. H.T, and F. L. wrote and revised the article including comments from all authors. All authors have read and approved the final version of the manuscript.

Funding

This work is supported by NIH Grants R01 MH111107 (FL & JZ), R01MH132226 (FL & JZ), and R43DA063261 (IonTX, Inc). J.Z. is also partly supported by the John D. Stobo, M.D. Distinguished Chair Endowment, and the Edith & Robert Zinn Chair Endowment in Drug Discovery.

Data availability

All data and code are available on GitHub: https://github.com/hamid-teimouri/Chemoinformatics_FGF14-Nav16_Hits_Analysis.

Declarations

Ethics approval and consent to participate

Not applicable.

Competing interests

F.L is the founder of IonTX Inc., a company dedicated to the commercialization of PPI modulators of Nav channels. T.J.B receives compensations as consultant at IonTX Inc. J.Z. receives funds as subaward PI from IonTX through NIH/NIDA 1R43DA063261.