Confusion with negative motifs

[PelFritz]

I have a binary classifier trained and computed deepLIFT scores for just the genomic sequences that belong to class 1. My expectation was to get only positive motifs but I also get negative motifs. How does one interpret this? I consider positive motifs as those with upward facing web logos and negative motifs as those with downward facing web logos. As additional information if needed, my model ends with an output layer with a single neuron and sigmoid activation.
Any explanation to this will be appreciated.
Thanks

[akundaje]

Negative importance scores indicate that certain bases are not preferred i.e. antagonistic towards the prediction output. The top negative motifs are definitely real since they often capture well known repressors (e.g. ZEB/SNAI motifs in fibroblasts for examples). Some times, we also observe motifs with positive scores also having negative scores. This is also often real biology where in different sequence contexts, motifs can have a positive or negative effect on the output. For very rare negative motifs (with very few seqlets), they are likely just some erroneous sign issues with the attribution scores. Thanks Anshul.

…

On Wed, May 31, 2023 at 7:29 AM Peleke Fritz ***@***.***> wrote: I have a binary classifier trained and computed deepLIFT scores for just the genomic sequences that belong to class 1. My expectation was to get only positive motifs but I also get negative motifs. How does one interpret this? I consider positive motifs as those with upward facing web logos and negative motifs as those with downward facing web logos. As additional information if needed, my model ends with an output layer with a single neuron and sigmoid activation. Any explanation to this will be appreciated. Thanks — Reply to this email directly, view it on GitHub <#109>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AABDWEOTCSB4LBWZT4FITJLXI5ITLANCNFSM6AAAAAAYVR2WXU> . You are receiving this because you are subscribed to this thread.Message ID: ***@***.***>

[PelFritz]

Thanks @akundaje, I think this makes sense with what we see. We could also see motifs appearing in both metaclusters (positive and negative) and suspected it was also as a result of the sequence context just like said.

Just for the sake of clarity, will it then be safe to attribute the negative motifs as those driving the decision towards the negative class (class 0) in a binary classification task?

Thanks
Fritz

[akundaje]

Yes correct. Negative motifs are favoring the negative class.

[PelFritz]

Thank you @akundaje for the quick responses.