Applying CHEESE Embeddings to Uncover Hidden Structural Patterns
How CHEESE Embeddings achieved a 67% hit recall and 1st place in the 9ADD Challenge, outperforming classical Morgan fingerprint-based approaches
Fingerprint-based methods for molecular description are a major cornerstone of computer-aided drug discovery and cheminformatics in general. However, due to their inherent limitations - including loss of three-dimensional information, bit collisions, and lossy structural approximation - they frequently underperform in workflows involving sterically or conformationally constrained binding sites.
In drug research especially, fingerprints suffer from their discrete, non-differentiable nature - a property that fundamentally limits their utility in modern generative and optimization workflows where gradient-based methods are increasingly employed. Furthermore, each substructural feature is encoded independently, without regard to its molecular context, which contributes to poor handling of activity cliffs and limits the ability to capture global molecular properties such as conformational flexibility and, in the case of electrostatic descriptors, electron density distribution.
In this case study, we apply CHEESE Embeddings as novel molecular descriptors that overcome the aforementioned limitations of classical fingerprint-based approaches by transforming each molecule into a 256-dimensional vector derived from its electrostatic or shape properties (EspSim or ShapeSim, respectively). As demonstrated below, these embeddings are capable of uncovering structural similarities between molecules in large datasets that remain undetected by the classical Morgan fingerprint-based approach.
Dual specificity tyrosine-phosphorylation-regulated kinase 1B (DYRK1B) is an enzyme proposed to be involved in the regulation of nuclear functions.[1] Its clinical significance stems from its paradoxical role as an oncogenic mediator: DYRK1B functions as a negative regulator of cell cycle progression while being markedly overexpressed in a variety of hyperproliferative malignancies.[2] This positions DYRK1B as a compelling drug target in the field of anticancer therapy development.
This case study was carried out as part of the 9ADD Challenge. The task was to identify 100 compounds with the highest potential as binders of DYRK1B from an organizer-provided blind set of 3,365 molecules. The training dataset consisted of 1,891 molecules, including 734 known actives and 1,157 confirmed inactives. Leveraging this dataset, a CHEESE Embeddings-based strategy was employed as follows:
Initially, a Morgan fingerprint-based t-SNE of the provided chemical space was computed. As illustrated in the diagram below, Morgan fingerprints rendered the entire dataset as a single, undifferentiated cluster with no meaningful separation. However, upon computing the t-SNE based on shape embeddings, the resulting diagram revealed a clear differentiation of the chemical space into two distinct regions: molecules dissimilar to the known actives, and compounds sharing a considerable degree of shape similarity based on the computed ShapeSim CHEESE Embeddings. This differentiation allowed us to rapidly exclude irrelevant molecules from further consideration.
In the next step, selecting the 100 nearest neighbors to the confirmed actives within the shape-similar region of the chemical space resulted in a 53% hit rate upon submission.
In the next attempt, we trained a LightGBM model on shape-based CHEESE Embeddings using the provided dataset. Application of this model to the blind set yielded a 33% hit rate.
Finally, we constructed a consensus of both approaches, combining results from the CHEESE Embeddings similarity search and the LightGBM model. This strategy resulted in a remarkable 67% hit rate, placing our team in 1st place in the 9ADD Challenge and substantially outperforming classical fingerprint-based methods.
The challenge is publicly available for those wishing to benchmark their own approaches - here.
The code used in our solution is provided in a public GitHub repository.
A recording of the solution presentation is available on YouTube (timestamp 37:10 - 50:50):