Back

Aggregating multimodal cancer data across unaligned embedding spaces maintains tumor of origin signal

Kirchgaessner, R.; Keutler, K.; Sivakumar, L.; Song, X.; Ellrott, K.

2025-05-18 bioinformatics
10.1101/2025.05.14.653900 bioRxiv
Show abstract

AI based embeddings offer the possibilities of encoding complex biological data into low dimensional spaces, called embedding spaces, that maintain the relationships between entities. There is an open question about the compatibility of embedding spaces that are created without any coordination. It has been assumed that signals in these unaligned embedding spaces would be destroyed if vectors were aggregated into summed values. We trained embedding models across different data modalities and tested aggregating the values together to test this assumption. Our research shows that signal from unaligned embedded values is conserved and able to still be used for learning tasks, such as data modality and tumor of origin recognition.

Matching journals

The top 7 journals account for 50% of the predicted probability mass.

50% of probability mass above

"Similar papers" are the closest papers from that journal in the model's embedding space. They show what the match is built on, but the ranking comes mostly from a classifier over the whole training set, not from these examples alone.