SPACE: Your Genomic Profile Predictor is a Powerful DNA Foundation Model

Jiwei Zhu, Zhao Yang, Bing Su

Track: long paper (up to 6 pages)

Keywords: DNA Foundation models, mixture of experts, genomic profile prediction, DNA, biology

Abstract: While unsupervised DNA pre-training has shown promise, we argue that supervised genomic profile prediction provides more effective DNA representations, since DNA functions are regulated by genomic profiles like chromatin accessibility. We propose **S**pecies-**P**rofile **A**daptive **C**ollaborative **E**xperts (SPACE), a model that uses Mixture of Experts (MoE) to capture cross-species and multi-profile relationships in genomic data. Through extensive evaluation, SPACE achieves state-of-the-art performance, demonstrating that supervised training with genomic profiles creates powerful DNA representations.

Submission Number: 3