Trait2Vec: Ontology-aware embeddings for organismal trait descriptions

Juan Jose Garcia, James Balhoff, Hilmar Lapp

Submission Track: Short papers presenting ongoing research or work submitted to other venues (up to 5 pages, excluding references)

Keywords: representation learning, ontology, trait descriptions, transformer

TL;DR: We propose Trait2Vec a machine learning model to embed organismal trait descriptions into a latent space where embedding similarity correlates with ontology-based metrics.

Abstract: Trait descriptions characterize how an organism looks, behaves or interacts. These descriptions are typically represented as text, but may be manually mapped within an ontology for downstream analysis. Nonetheless, the cost of this manual mapping is not scalable. In this work we propose a method to finetune a transformer model and embed textual trait descriptions in a latent space that captures the notion of distance within an ontology. The resulting model, which we coin Trait2Vec, can then embed trait descriptions in a scalable and biologically meaningful computational representation.

Submission Number: 56