Finding the Sweet Spot

Abstract

Complex carbohydrates, or glycans, are involved in cancer progression and could thereby serve as diagnostic markers as well as therapeutic targets. However, new research is required to determine which glycan motifs are universal signatures of tumor environments. Hence, in this project, we collected relative abundances of glycan structures in various tumor and healthy tissues from previous publications. Using this data we constructed two binary classification deep neural networks capable of predicting the health status (cancer or control) of a given glycan or glycan profile. Despite the complexity and diversity of data, both exhibit an accuracy of 80 % during validation and generalize rather well to test data. By extracting features important for model classification and through statistical analysis we could then identify sialic acid as being a prominent feature of tumor glycans and detect interesting changes in the role of mannose depending on glycan type. Even more motifs could be of potential interest, though these results need to be further substantiated. We envision that these contributions could serve as a stepping stone for future research in the field, and see potential for development e.g., through construction of a multiclass model predicting cancer types.