ML models that use SBOL for DNA annotation

[Gonza10V]

Background

Machine Learning (ML) quickly advanced in the last few years. Key to the ML models is the training process where models learn from data. Therefore, data quality and quantity are fundamental to model performance. Datasets like ImageNet have promoted the use of ML on the Computer Vision field for image classification leading to major advances like the development of AlexNet and ResNet for example. Synthetic Biology (SynBio) looks for engineering biological systems and has created abstractions and standards for that endeavor. Sequence-to-expression is a hallmark for SynBio but the field lacks easy to use datasets so researchers and developers can focus on creating new models instead of gathering and preprocessing data. In a previous project we started the development of SeqTrainer to create ML datasets from SBOL [1] designs SynBioHub. To advance this endeavor we need to include tokenizers and ML models tailored for DNA sequence data.
In this project we will explore the use of foundational models such as Evo 2, DNABert, HyenaDNA and TITANS to predict constitutive expression and label DNA with sequence features from SBOL such as promoters and coding sequences.

[1] Buecherl, Lukas, et al. "Synthetic biology open language (SBOL) version 3.1. 0." Journal of integrative bioinformatics 20.1 (2023): 20220058.
[2] Urtecho, Guillaume, et al. "Genome-wide Functional Characterization of Escherichia coli Promoters and Sequence Elements Encoding Their Regulation." eLife 12 (2023).

Goal

Develop models to for DNA annotation using SBOL

Specific Goals:
Develop workflows to train models using HPC.
Explore baseline performance of existing models.
Modify a model, domain adapt and finetune for bacteria.
Document the developments and create example notebooks.

Difficulty Level: Medium

This project involves applying ML models for predicting expression, DNA annotation using SBOL and its inclusion in a Python package for data set creation and training on ML models.

Size and Length of Project

• medium: 175 hours
• 16 weeks

Skills

Essential skills: Python, ML, GitHub, Git
Nice to have skills: SBOL, LLM

Public Repository

https://github.com/SynBioDex/SeqTrainer

Potential Mentors

Gonzalo Vidal (Gonzalo.vidalpena@colorado.edu)
Chris Myers

[qzMalekuz]

Hi @Gonza10V and maintainers,
I’ve started exploring this issue and have set up the repository locally I’m currently working on parsing SBOL files and understanding how sequences are extracted for ML pipelines.

I plan to:

• Build a small SBOL → dataset pipeline
• Test a simple baseline model for annotation

Could you suggest:

• Recommended datasets?
• Preferred model direction (DNABERT vs simpler baselines)?

Looking forward to contributing!

Thanks,
Zafar.

[Gonza10V]

Hi @qzMalekuz, the dataset is too big to store it on GitHub so you have to download it. Check the readme of SeqTrainer for further instructions. We already have simpler models like CNNs that replicate results from scientific publications. So the model direction is towards more complex models such as DNABERT, here is the repo that I suggest https://github.com/MAGICS-LAB/DNABERT_2 , it has an example for finetuning and you can use this dataset https://github.com/SynBioDex/SeqTrainer/blob/main/data/data_DNABERT/promoter_classification_DNABERT.zip . We then want to move into testing Evo 2 (https://github.com/ArcInstitute/evo2).

[qzMalekuz]

Hi @Gonza10V , thanks a lot for the detailed guidance!

I’ll start by setting up SeqTrainer locally and running the baseline models using the provided dataset to understand the full pipeline.

After that, I plan to focus on building a bridge from SBOL data to the dataset format used for training, so that SBOL-based inputs can directly flow into the existing pipeline.

I’ll keep you updated with progress and share a PR soon.

Thanks,
Zafar.

[qzMalekuz]

Hi @Gonza10V , I’ve opened a PR (SynBioDex/SeqTrainer#5 (comment)) adding an initial SBOL → CSV dataset conversion script that extracts sequences and resolves references from SBOL documents.

It’s designed as a first step toward integrating SBOL-based inputs into the SeqTrainer pipeline.

Would really appreciate your feedback on whether this aligns with the expected direction!

thanks,
Zafar.

[sreyacv1]

Hi @Gonza10V, I am a bioinformatics student focused on the bacterial domain adaptation and fine-tuning parts of this project.
I am currently setting up SeqTrainer and reviewing the DNABERT_2 and Evo 2 architectures. Since my background is in genomics, I can help ensure the models correctly capture prokaryotic motifs like -10/-35 boxes during training.
Quick question: For the bacterial domain adaptation, are we focusing strictly on E. coli datasets, or should the pipeline handle diverse taxa?

[Gonza10V]

I would focus the domain adaptation on gram negative, but I'm open to try different aproaches depending on results. More focus to E. coli and wider domain adaptation including all bacteria are posibilities that we could explore.

[Saurabhtbj1201]

Hello @Gonza10V , I am a GSoC 2026 applicant interested in working on this project.
I have experience with Python and machine learning, and I am currently exploring the SeqTrainer repository and SBOL-based datasets to better understand the workflow. I am particularly interested in working on dataset preparation, tokenizer development, and evaluating baseline models such as DNABERT.
I would appreciate any guidance on how to get started, especially regarding recommended initial tasks or areas where I can contribute effectively. I would also be happy to take on a small task or issue to begin contributing.
Thank you!