Welcome to the AI Hackathon on Genomic Language Model hosted by the StatBiomed lab at the University of Hong Kong.

Materials are stored in this GitHub repo: https://github.com/StatBiomed/DNA-AI-hackathon

Schedule

  • Full-day intense hackathon: April 2nd (8.30am-8.30pm) briefing at Seminar Rm 1A, G/F, 5 Sassoon Rd, 2025

  • Tutorial: March 26th, 2025, 2-3.30pm at Seminar Rm 1A, G/F, 5 Sassoon Rd, by Dr Shumin Li (& Fangxin Cai / Ruiyan Hou)

  • Seminar: March 25th, 2025, 4-5pm at Lecture Theatre 1, 21 Sassoon Rd, by Dr Shumin Li

Candidate models

  1. Evo2 (bioRxiv): 7B or 40B with trained on 9.3 trillion DNA base pairs (in 1 million token contexts)

  2. NucleotideTransformer (paper) Model: ranging from 50 million up to 2.5 billion parameters and integrating information from 3,202 human genomes and 850 genomes from diverse species

  3. Caduceus (arXiv) MambaDNA-based model

  4. GET (paper) general expression transformer

Tasks

Task1: SNP2GEX

See details in ./task1_SNP2GEX/

In this task, we will use individuals’s genomic sequence (with both common and rare variants) to predict its cis gene expression, at both gene level expression (task 1a) or promoter activities (task 1b).

The success of this task will have a huge impact on precision medicine, including the study of the functional effects of rare variants and somatic variants, which lack power in common eQTL studies.

Multiple labs have been working on this, but it remains an extremely challenging task. We are benchmarking multiple state-of-the-art genomic language models.

Task2: Seq2CellxTF

See details ./task2_Seq2CellxTF/

In this task, we will use the reference sequence and its regulatory regions to predict the cell type-specific transcription factor bindings.

As you can see, the input data is limited, mainly the relatively short sequence, optionally with its cis-gene expression (maybe chromatin open accessibility in the future), but the output is extremely large, not only for many TFs but also in diverse cell conditions.

This would be an ideal task to demonstrate the zero-shot capability of “foundation model”

Sponsors

  • HKU’s CPOS for the GPU support!
  • HKUMed & School of Biomedical Sciences