It sometimes happens that, in the raw data, a protein has multiple annotations linking it to the same GO tag but with conflicting qualifiers. Protein A0A087X1C5 is an example: it has two listings for GO:0070330, one with the \"enables\" qualifier, and one with the \"NOT|enables\" qualifier. Both list \"IDA\" as the evidence type. Following the links on the QuickGO page above, I see that the former comes from this 2004 paper by Pai et al., while the latter comes from this 2008 paper by Zhang et al. I'll admit that my reading comprehension for these two papers is pretty low, but on a quick scan, the Zhang et al. paper seems to me to be, at least in part, a direct rebuttal of the Pai et al paper. Which is fine; one should expect this to happen in the literature from time to time.
\nMy understanding is that the procedure for handling cases like this in the contest labels is to discard the listing with the \"NOT\" qualifier, so that the GO:0070330 annotation appears in train_terms.tsv. This has the practical advantage that the procedure is clear and easy to implement. I can imagine a variety of other reasons why this might be the best approach for this contest. For example: perhaps conflicts like this are negligibly rare (I haven't checked this yet); perhaps it is more difficult to convincingly establish experimentally that a protein does not have some function than it is to establish that it does; etc. But coming from me, arguments like that are just speculation. So I'm wondering if any subject matter experts in this forum might be willing to weigh in: in cases where there are annotations with conflicting qualifiers and equal evidence codes, why do we discard the one with the \"NOT\" qualifier, rather than, say, taking the more recent publication or somehow modifying the prediction target and loss to reflect that both are present?
\n(…asking because I'm thinking through whether and how I want to attempt to build an analysis that uses the NOT annotations and/or the annotations supported only by weaker evidence types somehow.)
", "messages": [ { "id": 3310430, "postDate": "2025-11-02T22:15:53.457Z", "content": "It sometimes happens that, in the raw data, a protein has multiple annotations linking it to the same GO tag but with conflicting qualifiers. Protein A0A087X1C5 is an example: it has two listings for GO:0070330, one with the \"enables\" qualifier, and one with the \"NOT|enables\" qualifier. Both list \"IDA\" as the evidence type. Following the links on the QuickGO page above, I see that the former comes from this 2004 paper by Pai et al., while the latter comes from this 2008 paper by Zhang et al. I'll admit that my reading comprehension for these two papers is pretty low, but on a quick scan, the Zhang et al. paper seems to me to be, at least in part, a direct rebuttal of the Pai et al paper. Which is fine; one should expect this to happen in the literature from time to time.
\nMy understanding is that the procedure for handling cases like this in the contest labels is to discard the listing with the \"NOT\" qualifier, so that the GO:0070330 annotation appears in train_terms.tsv. This has the practical advantage that the procedure is clear and easy to implement. I can imagine a variety of other reasons why this might be the best approach for this contest. For example: perhaps conflicts like this are negligibly rare (I haven't checked this yet); perhaps it is more difficult to convincingly establish experimentally that a protein does not have some function than it is to establish that it does; etc. But coming from me, arguments like that are just speculation. So I'm wondering if any subject matter experts in this forum might be willing to weigh in: in cases where there are annotations with conflicting qualifiers and equal evidence codes, why do we discard the one with the \"NOT\" qualifier, rather than, say, taking the more recent publication or somehow modifying the prediction target and loss to reflect that both are present?
\n(…asking because I'm thinking through whether and how I want to attempt to build an analysis that uses the NOT annotations and/or the annotations supported only by weaker evidence types somehow.)
", "rawMarkdown": "It sometimes happens that, in the raw data, a protein has multiple annotations linking it to the same GO tag but with conflicting qualifiers. Protein [A0A087X1C5](https://www.ebi.ac.uk/QuickGO/annotations?geneProductId=A0A087X1C5) is an example: it has two listings for GO:0070330, one with the \"enables\" qualifier, and one with the \"NOT|enables\" qualifier. Both list \"IDA\" as the evidence type. Following the links on the QuickGO page above, I see that the former comes from [this 2004 paper by Pai et al.](https://www.jbc.org/article/S0021-9258(20)85400-3/fulltext), while the latter comes from [this 2008 paper by Zhang et al.](https://europepmc.org/article/MED/18838503) I'll admit that my reading comprehension for these two papers is pretty low, but on a quick scan, the Zhang et al. paper seems to me to be, at least in part, a direct rebuttal of the Pai et al paper. Which is fine; one should expect this to happen in the literature from time to time.\n\nMy understanding is that the procedure for handling cases like this in the contest labels is to discard the listing with the \"NOT\" qualifier, so that the GO:0070330 annotation appears in train_terms.tsv. This has the practical advantage that the procedure is clear and easy to implement. I can imagine a variety of other reasons why this might be the best approach for this contest. For example: perhaps conflicts like this are negligibly rare (I haven't checked this yet); perhaps it is more difficult to convincingly establish experimentally that a protein does not have some function than it is to establish that it does; etc. But coming from me, arguments like that are just speculation. So I'm wondering if any subject matter experts in this forum might be willing to weigh in: in cases where there are annotations with conflicting qualifiers and equal evidence codes, why do we discard the one with the \"NOT\" qualifier, rather than, say, taking the more recent publication or somehow modifying the prediction target and loss to reflect that both are present?\n\n(...asking because I'm thinking through whether and how I want to attempt to build an analysis that uses the NOT annotations and/or the annotations supported only by weaker evidence types somehow.)", "votes": 1 } ], "comments": [], "raw_markdown_by_id": { "3310430": "It sometimes happens that, in the raw data, a protein has multiple annotations linking it to the same GO tag but with conflicting qualifiers. Protein [A0A087X1C5](https://www.ebi.ac.uk/QuickGO/annotations?geneProductId=A0A087X1C5) is an example: it has two listings for GO:0070330, one with the \"enables\" qualifier, and one with the \"NOT|enables\" qualifier. Both list \"IDA\" as the evidence type. Following the links on the QuickGO page above, I see that the former comes from [this 2004 paper by Pai et al.](https://www.jbc.org/article/S0021-9258(20)85400-3/fulltext), while the latter comes from [this 2008 paper by Zhang et al.](https://europepmc.org/article/MED/18838503) I'll admit that my reading comprehension for these two papers is pretty low, but on a quick scan, the Zhang et al. paper seems to me to be, at least in part, a direct rebuttal of the Pai et al paper. Which is fine; one should expect this to happen in the literature from time to time.\n\nMy understanding is that the procedure for handling cases like this in the contest labels is to discard the listing with the \"NOT\" qualifier, so that the GO:0070330 annotation appears in train_terms.tsv. This has the practical advantage that the procedure is clear and easy to implement. I can imagine a variety of other reasons why this might be the best approach for this contest. For example: perhaps conflicts like this are negligibly rare (I haven't checked this yet); perhaps it is more difficult to convincingly establish experimentally that a protein does not have some function than it is to establish that it does; etc. But coming from me, arguments like that are just speculation. So I'm wondering if any subject matter experts in this forum might be willing to weigh in: in cases where there are annotations with conflicting qualifiers and equal evidence codes, why do we discard the one with the \"NOT\" qualifier, rather than, say, taking the more recent publication or somehow modifying the prediction target and loss to reflect that both are present?\n\n(...asking because I'm thinking through whether and how I want to attempt to build an analysis that uses the NOT annotations and/or the annotations supported only by weaker evidence types somehow.)" } }