Hello, I want to compare AlphaGenome against experimental data from various papers. One of these comparisons is MyC binding across 12 different cell lines. So far, I’ve noticed that only 6 of those cell lines are available in AlphaGenome, and of those 6, only 2 have any TF chIP-seq predictions for MyC. This is a result of filtering output = model.predict_interval(…) by its transcription factor metadata, i.e.
output_chip_tf = output.chip_tf.filter_tracks(
(output.chip_tf.metadata['transcription_factor'].isin(['MYC'])).values
)
Is this the correct approach for determining which cell lines/biosamples have a specific chIP TF track? If it is, how might I be able to figure out what cell lines have what chIP TF tracks without having to make the predictions?
As a separate question, in general, is there any way to have AlphaGenome make predictions for biosamples that aren’t immediately available (for instance, any predictions for snu398, or MyC chIP TF predictions for MM.1S)?
Hi There!
Thanks for reaching out.
Yes, that’s correct. AlphaGenome’s outputs are tied to the experimental datasets it was trained on so not all modalities will be available for all biosamples. We have the following navigating data ontologies Colab notebook that has a table of all our metadata, and you can programmatically interact with this data using an API key, or indeed download as a CSV. This will allow you to explore what modality/biosamples pairings are available.
You can search for biosample name or specific ontology identifiers. If the track you want isnt available, you can try looking for proxy tissues or cell lines. If tracks for similar tissue/cell types aren’t available, you could look for high-consensus variants (those that show a strong effect across most cell types that are similar). If a variant has a high effect across 90% of tracks, it is likely a constitutive regulatory element and should show up in your primary cell line. As part of a cell-type agnostic approach, you could extract variant effect predictions across a set of cell types and modalities, then use a model like LASSO regression to aggregate these features and train it against your experimental data.
Kind regards,
Tumi