Questions About Using TACIT for Single-Species Model Training with Multi-Species MAF

[aaannaw]

Dear Author,

I hope this email finds you well. I am reaching out because I am interested in using TACIT to train a single-species model inspired by the approach in your paper, "Relating enhancer genetic variation across mammals to complex phenotypes using machine learning," particularly the MouseMotorCortexModel.

My Data and Goals
Multi-species MAF file: I have a whole-genome alignment (MAF format) for 12 rodent species.

ATAC-seq data: For only one of these species, I have ATAC-seq peaks from three tissues.

Objective: Train a tissue-specific model (similar to MouseMotorCortexModel) to predict regulatory elements and link them to phenotypes.

Questions
After reviewing the TACIT Guidelines and GitHub documentation, I still have a few uncertainties:

ATAC Data Alignment for Open Chromatin Regions

Should I map my ATAC-seq reads only to my target species’ genome, or should I align them to all species in the MAF?

Generating Positive/Negative Sets

Once I define open chromatin regions, are there TACIT scripts to help generate positive (enhancers) and negative (non-enhancers) sets?

Model Training with Single-Species Data

The GitHub documentation suggests starting with ocr_phylo(g)lm.r to test OCR-phenotype associations, but I only have ATAC peaks from one species. Can I still train a model without phylogenetic correction?

If so, what would be the recommended workflow? (e.g., train a baseline model first, then refine with cross-species data later?)

My ultimate goal is to link predicted enhancers to phenotypes, but I’m unsure how to proceed with only single-species ATAC data.

I would greatly appreciate any guidance or suggestions you might have. Thank you for your time and for developing this fantastic tool!

Best regards,
Na Wan

[imk1]

@aaannaw I am so sorry that I did not notice this until now. Here is some feedback:

1. With only 12 species in your alignment, you probably do not have enough power for TACIT. If any of those species are in a larger alignment (https://cglgenomics.ucsc.edu/data/cactus/), I would recommend using the rodents from the larger alignment too.

2. If you have data from only one species, then you cannot evaluate how well your model can predict changes in open chromatin status between species. I have found that not all models that work well on part of a genome not used in training can predict open chromatin status differences between species (https://www.youtube.com/watch?v=RKbHrIGdHrw). I generally recommend starting with training a model in one species and evaluating the model on species not used in training to ensure that it is possible to train a model that learns a regulatory code that is shared across species. If that works well, then I re-train the model using the data from all of the species. If you have data from 1-2 tissues in additional species, I would focus on those tissues. If you have data from only 1 species other than house mouse for all of your tissues but there is public data from house mouse (check out https://www.encodeproject.org/matrix/?type=Experiment&control_type!=*&status=released&perturbed=false&assay_title=ATAC-seq&replicates.library.biosample.donor.organism.scientific_name=Mus+musculus and https://www.nature.com/articles/s41597-019-0071-0), I would evaluate your model using the data from house mouse and then integrate that data into training. If there is no way to get data from an additional species, I recommend not doing TACIT because you cannot evaluate whether your model can predict changes in open chromatin status between species.

3. I recommend mapping your reads to the genome of the species from which they were generated. I would then call peaks and use multi-species alignments to map the peaks to other species.

4. We do not have exact scripts for generating positive and negative sets because exactly how you generate them depends on the data you have. For example, we defined "enhancers" to be reproducible peaks at least 20kb from a transcription start site, but, if you are working with a smaller genome, than using peaks closer to the transcription start site might be o.k. For negatives, if you have data from enough species, you can use closed chromatin regions whose orthologs in another species for which you have data are open, but, if you have data from fewer species, you may need to add G/C-matched regions and, if you have them, peaks in a similar tissue.

5. The phylogenetic correction is not for model training but rather for associations with phenotypes. After you train the model, you can get the regions in other species corresponding to your peaks, make predictions at those regions, and then associate the predictions across species with your phenotype.

I appologize again for the delayed response. Feel free to post additional questions.