The labels in Gene Function Classification are not Mutually Exclusive

#19
by Linear-Matrix-Probability - opened

Thank you for your excellent contributions! I'm delving into the realm of gene function classification and have noticed that the labeling system employed doesn't seem to adhere to mutually exclusive categories. For instance, the gene named ENSG00000285708 is referenced across three distinct files. Would it be accurate to approach this task as a multi-label classification problem?

Thank you for your interest in Geneformer! Geneformer is a foundation model that can be fine-tuned towards a multitude of downstream tasks. For each task, genes can have different labels. For example, in an analogy to NLP, if you were classifying parts of speech, the word “nice” would have the label “adjective”. If you were classifying by sentiment, the word “nice” could have the label “positive”. If you wanted to train a model to jointly understand parts of speech and sentiment you could do that, or if you’d like to train a model specifically to understand parts of speech, you would only use the labels relevant to parts of speech.

The above assumes you are asking why genes are contained in multiple task-specific datasets. If you have a different question, please specify which distinct files you are referring to. Thank you!

ctheodoris changed discussion status to closed

Thank you for your prompt reply. I'm very sorry that I didn't specify the corresponding file. The gene with the number ENSG000000285708 appears in all three files you provided: bivalent_gene_labels.txt, lys4_only_gene_labels.txt, and no_methylation_gene_labels.txt. Does it indicate this gene have all three labels simultaneously: bivalent, non-methylated, and lys4-only methylated? In your article, two binary classification tasks related to these labels were performed, namely Bivalent versus Non-methylated, Bivalent versus Lys4-only methylated. So, I feel a bit confused.

Thank you for clarifying what files you were referring to. The chromatin marks are recorded per known TSS in Bernstein et al, Cell 2006 within the 56 highly conserved loci. Certain genes appear in multiple groups (see Fig. 2 of that reference), likely due to alternative TSS's. Therefore, biologically, to answer your original question, this is not a multi-label classification problem. Since these are just meant to teach the model about patterns of bivalently marked genes and it's not critical that all genes are included, it may be most useful for fine-tuning to exclude genes that appear in multiple classes.

Thank you very much for your response. My confusion has been resolved completely.

Sign up or log in to comment