How well do genomic sequence-to-function models generalize to non-model animal genomes?

Purpose

The aim of this project is to evaluate how well sequence-based machine-learning models for functional genomics generalize beyond the species on which they were developed and trained.

The project will benchmark predictions from models such as AlphaGenome against experimentally measured functional genomics data from non-model animal species, for example livestock, aquaculture, sport and companion animals.

Background

A large proportion of biologically important variation is located outside protein-coding genes. These non-coding regions can regulate when, where and how strongly genes are expressed and therefore contribute substantially to variation in complex traits.

Recent deep-learning models can predict functional genomic features directly from DNA sequence. These models can, for example, predict signals related to chromatin accessibility, gene expression and regulatory activity over large genomic regions.

One recent example is AlphaGenome (Avsec et al. 2026), developed by Google DeepMind, which predicts a wide range of genomic signals directly from DNA sequence. However, such models are predominantly developed and evaluated using data from humans and a limited number of model organisms.

It therefore remains unclear how well their predictions transfer to more evolutionarily distant species. This is particularly relevant for livestock, aquaculture, sport and companion animals, where functional genomics datasets are becoming increasingly available but are still much less extensive than in humans and traditional model species.

Project description

This project will use publicly available functional genomics datasets to benchmark sequence-to-function models across different animal species.

The student will identify suitable datasets, process experimentally observed genomic signals and compare these with model predictions generated from the corresponding DNA sequences.

Depending on the student's interests and the availability of suitable datasets, the project may include:

• Collecting and curating publicly available functional genomics datasets, such as ATAC-seq and RNA-seq.
• Identifying suitable reference genomes and genomic annotations for the selected species.
• Applying sequence-based genomic prediction models to non-model animal genomes.
• Comparing predicted genomic signals with experimentally observed data.
• Evaluating predictive performance using appropriate statistical metrics.
• Comparing performance across species.

The exact scope can be adapted depending on the student's interests, computational experience and the availability of high-quality public datasets.

Specifications

Suitable for: MSc students in Bioinformatics, Animal Science, Genetics, Biology, Data Science or related fields.

Experience with programming in Python or R is helpful. Basic familiarity with genomics or sequence data is an advantage, but previous experience with deep learning is not required.

The project is particularly suitable for students interested in genomics, bioinformatic benchmarking and comparative biology.

The student will gain experience with:

• Functional genomics data.
• Sequence-based prediction models.
• Processing and analysis of public genomic datasets.
• Machine-learning model evaluation.