| Tartalmi leírás: |
Efficiently predicting interactions between compounds and target proteins is pivotal in drug discovery, driving the need for machine-learning-based approaches to replace resource-intensive experiments. This study investigates the potential of hyperbolic geometry in enhancing pairwise interaction prediction models, emphasizing the identification of suitable input modalities to leverage additional information through nonEuclidean embeddings. Despite the potential benefits of hyperbolic embeddings, our study highlights their limited contribution when employing widely used structure-based pre-trained input representations. However, hyperbolic predictors outperform their Euclidean counterparts with transcriptomics-based input, underscoring the importance of an appropriate input modality, such as differentially expressed gene signatures. Besides aiding the selection of input modalities for interaction prediction, the results also confirm our prior hypothesis: differential expression signatures possess a non-Euclidean nature and thus can be better represented in a hyperbolic vector space.
|
