3D Molecular Analysis with CHEESE Embeddings: A Neurosteroid Case Study
Since their discovery in 1981 [1], steroid hormones known as neurosteroids, have shown a crucial role in a plethora of neuronal functions, such as cognition, memory, neuroprotection or myelination [2]. Their mechanism of action mostly revolves around nervous system modulation of neuron ionic channels and receptors. A specific class of neurosteroids targets are GABAA , NMDA receptors, voltage-dependent T-type Ca2+ or voltage-dependent K+ channels, receptors and channels involved in pathways related to anesthesia and analgesia modulation [3]. Neurosteroids bearing such antinociceptive features - inhibiting pain detected by sensory neurons - are the cornerstone of this study.
Given the typical structure and shape of steroidal compounds, together with their sensitivity towards even a minor structural changes from the perspective of the target affinity, computational approach towards the activity prediction brings a great advantage. As we reported previously [4], CHEESE Embeddings, combined with ML-techniques, are a powerful tool for analysis and elucidation of even subtle structural and electrostatic nuances. In this case CHEESE Embeddings were applied in the methodology enabling activity prediction of novel tetrazolic neurosteroids before they ever enter the lab - specifically, compounds expected to act as positive allosteric modulators (PAMs) of GABAA receptors and deliver the desired antinociceptive effect.
Studied Candidates and Training Set
A series of studied tetrazole-functionalized neurosteroids consisted of six unique structures sharing a tetrazole ring attached at position C-17. Their stereochemistry varied at positions C-3 and C-5, which introduced a significant conformational diversity to the studied dataset (see Figure 1).
As a training set, a library of 124 active and inactive neurosteroids was employed, sourced from the available literature. The set was curated to ensure homogeneity of the underlying bioactivity data.
To evaluate the structural diversity of the training set, a pairwise 3D shape similarity matrix was computed for all 124 compounds.
The resulting heat map reveals four distinct blocks of structural similarity along the diagonal, suggesting that the library naturally stratifies into structurally coherent groups.
3D Shape Similarity Analysis
Employing Shape Similarity CHEESE Embeddings for the study of molecular shape features, k-medoids clustering (k=4) partitioned the dataset into four structurally distinct groups. Each cluster was thus represented by a medoid - a molecule reflecting the structural center of the cluster and serving as its most representative member.
In order to identify a discriminative medoid, a scoring function defined as follows was applied to each cluster medoid, rewarding high shape similarity to active compounds while penalizing similarity to inactive ones:
where Sij denotes the shape similarity between compound i and active compound j, Sik the similarity to inactive compound k, and λ a weighting parameter balancing the two terms. The medoid yielding the highest score was designated as the discriminative medoid - the molecule whose 3D shape best separates actives from inactives across the entire dataset. Subsequent visualization of the 3D shape-based chemical space of all 124 training molecules and 6 tetrazolic neurosteroids is shown in Figure 3 below:
A visual comparison of medoids from each cluster revealed an interesting finding - medoid 77 (cluster K2), although sharing the core structure with the studied compounds, is not a steroidal compound. Notably, clusters K1 and K2 exhibit mirror symmetry. Figure 4 below shows the medoid of each cluster together with the discriminative medoid - the optimal shape template identified across all clusters.
The discriminative medoid thus served as a shape template against which the six novel tetrazolic neurosteroids were scored, enabling activity prediction prior to synthesis. These predictions are discussed in detail in the following section alongside the electrostatic analysis.
3D Electrostatic Similarity Analysis
Electrostatic similarity CHEESE Embeddings were applied in a similar fashion as their shape similarity counterparts. Upon k-medoids clustering, four cluster medoids were obtained and the discriminative medoid was identified using the scoring function described above. The complete chemical space of the 3D electrostatic similarity analysis is shown in Figure 5.
Using the discriminative scoring function, compound 10 was identified as the discriminative medoid. Its structure is shown in Figure 6 below:
Predicting Activity: Shape, Electrostatic or Consensus?
All six tetrazolic neurosteroids were scored against both the 3D shape-based discriminative medoid 94 and the 3D electrostatic-based discriminative medoid 10, resulting in a score ranging from 0 to 1 for each compound and each method. Complete computed and experimental results are shown in Table 1 below.
Notably, the most active in vitro molecule NS1 was correctly ranked by the electrostatic-based model, reaching a high score of 0.931. However, this was not the case for the shape-based model, which predicted the compound to be inactive. A correct consensus appeared with NS2, which was predicted inactive by both models - a prediction later confirmed by experiment. An interesting conflict arose in the case of NS3 - predicted active by the Shape-based model but inactive by the electrostatic similarity-based model. Here, the electrostatic similarity model proved correct, as NS3 was experimentally found to be inactive. In the case of NS4, both models predicted the compound to be active. Unfortunately, experimental validation proved the compound to be inactive. Compound NS5 was predicted inactive by the Shape-based model and borderline by the electrostatic similarity-based model - a consensus leaning towards inactivity that was confirmed experimentally, as the compound showed no measurable PAM effect. Perhaps the most interesting case overall was NS6, which exerted a partial inhibitory activity when tested at 100 nM, suggesting weak but measurable GABAA PAM activity. This borderline outcome was neither clearly captured by the Shape-based nor the electrostatic similarity-based model, both of which predicted the compound as active.
Compound NS1 was subsequently advanced to in vivo evaluation in a rat model of osteoarthritis-induced pain. Administered intraperitoneally at 3 mg/kg in cyclodextrin, it demonstrated significant reduction in mechanical hypersensitivity comparable to FDA-approved reference compounds zuranolone and ganaxolone - validating the computational predictions made prior to synthesis.
This study demonstrated that 3D molecular similarity analysis using CHEESE Embeddings can reliably guide activity prediction in a structurally sensitive compound series such as neurosteroids. The electrostatic similarity-based model proved to be a more robust predictor of GABAA PAM activity than the Shape-based counterpart - correctly identifying NS1 as the top candidate despite its unfavorable shape score, and accurately flagging several inactive compounds. Shape similarity alone, while informative for understanding structural diversity and cluster organization, was insufficient as a standalone activity predictor in this dataset.
The results confirm that CHEESE Embeddings capture pharmacologically relevant molecular features - particularly electrostatic surface properties - that translate directly into binding activity. The near-perfect match between the electrostatic similarity-based prediction and the experimental EC50 of NS1 (score 0.931, EC50 = 325 nM) illustrates the practical value of embedding-based similarity for candidate prioritization.
Across six novel compounds, the electrostatic similarity-based model correctly classified five - a promising hit rate for a prospective, fully unsupervised prediction made prior to any synthesis. Noteworthy, NS1 achieved a high score 0.931 from the electrostatic similarity-based model alongside excellent biological activity of EC50 = 325 nM, further validated in vivo against FDA-approved zuranolone and ganaxolone. In this use case, electrostatic similarity CHEESE Embeddings therefore proved more effective at capturing the pharmacologically relevant features of structurally sensitive compound classes such as neurosteroids.
Are you interested in applying CHEESE Embeddings in your own research pipeline? Contact us via email or LinkedIn.
Or try out our CHEESE platform and search for enumerated and non-enumerated trillion-scale chemical spaces yourself!
Acknowledgement
The computational part was performed by Miroslav Lžičař, CTO at Deep MedChem. The synthetic part and experimental validation were performed in the research group of Dr. Eva Kudová at IOCB Prague.