I just wanted to share a brief post here for the computer scientists about the task at hand who may lack an understanding of what is being asked of them. To start with we need some defintions of things you may think are farily obvious but are actually not. For starters what is a gene, what is a protien, and what are their functions? These are critical since the predictive task is to map amino acid sequence of a protien to its \"funciton\" i.e Gene Ontology terms. We need to get into the weeds here. The description in the data seciton, https://www.kaggle.com/competitions/cafa-6-protein-function-prediction/data left several important things out we need to go over.
\nBasically you have DNA, composed of the 4 nucleotide bases which have a triplet code that gets read out into mRNA then into protien amino acid sequences. Here we use the amino acide sequences of the protien as input. Simple enough right? Except we need to consider where this data is coming from. Thousands of different experiments in differnt biological contexts, many different cell types in many different disease states in many different species and cell lines. The protiens are not unique to these contexts nor are the sequences.
\nWhat I want you to consider is that it is well known that humans share 99% of our coding protien genes with chimpanzes our closest genetic relative. What is less widely known is that we also share ~80% with a bananna or for that matter nearly any complex mutlicellular life and a good deal of the single cellular life realted to mulicellular life, on the eukaryotic branch of our family tree. Once nature evolves a gene it does not like to remake that gene. Instead it gets repurposed, and there are different levels to what we as a human might term a \"purpose\" or \"function\". This brings us to the gene ontology.
\nThe Gene Ontology organizes protein function into three axes: Molecular Function (what the protein does chemically), Biological Process (what larger goal it contributes to), and Cellular Component (where in the cell it operates). The same sequence can perform the same chemical activity in a completely different location while contributing to an entirely different process i.e the label is not a stable property of the sequence alone. On top of this, the output space is not a flat set of classes but a graph with tens of thousands of nodes, most proteins have very few labels, and crucially, a missing label does not mean \"no\" it means \"untested.\" You are being asked to learn a mapping where the inputs are the same across contexts, the outputs depend on context you don't have, and the training labels are incomplete by construction.
\nThere are a few terms about protiens I encourage you to look up so you understand why this is going to be really really hard. Mutlifunctionality aka \"moonlighting\", protiens serve multiple functions on many levels. For one thing the same amino acid sequence that allows it to perform a chemical reaction may actually be able to perform a slightly similar chemical rection and because biology is messy, it will happen. This means the sequence will not have a 1 to 1 MF mapping. Further more the molecular provcess itself appears in many different biological processes BP, and the protien in question will do this same MF for many differnt BPs as well as potentially a few different MFs in the same BP in siffernt contexts. Lastly you'd think the cellular component CC, where this is happening would be stable, but again its not. Protiens get shuffled around and end up in different locations.
\nI do not want to make it seem like there is no stability or predictability at all, its just that the really interesting biology does not happen here. The shared genes I mentioned earlier, they are the ones that are common across all biological contexts because when those get disrupted, the cell dies. Evolution works by repurposing what exists and this is true in disease states as well, like cancer.
\nMaybe some one here more intelligent than me can take this into account some how, just wanted to share my thoughts and point out its not simply a matter of applying the fanciest model here. You'll need to apply some domain expertise here as well.
", "messages": [ { "id": 3419870, "postDate": "2026-03-11T20:54:52.340Z", "content": "I just wanted to share a brief post here for the computer scientists about the task at hand who may lack an understanding of what is being asked of them. To start with we need some defintions of things you may think are farily obvious but are actually not. For starters what is a gene, what is a protien, and what are their functions? These are critical since the predictive task is to map amino acid sequence of a protien to its \"funciton\" i.e Gene Ontology terms. We need to get into the weeds here. The description in the data seciton, https://www.kaggle.com/competitions/cafa-6-protein-function-prediction/data left several important things out we need to go over.
\nBasically you have DNA, composed of the 4 nucleotide bases which have a triplet code that gets read out into mRNA then into protien amino acid sequences. Here we use the amino acide sequences of the protien as input. Simple enough right? Except we need to consider where this data is coming from. Thousands of different experiments in differnt biological contexts, many different cell types in many different disease states in many different species and cell lines. The protiens are not unique to these contexts nor are the sequences.
\nWhat I want you to consider is that it is well known that humans share 99% of our coding protien genes with chimpanzes our closest genetic relative. What is less widely known is that we also share ~80% with a bananna or for that matter nearly any complex mutlicellular life and a good deal of the single cellular life realted to mulicellular life, on the eukaryotic branch of our family tree. Once nature evolves a gene it does not like to remake that gene. Instead it gets repurposed, and there are different levels to what we as a human might term a \"purpose\" or \"function\". This brings us to the gene ontology.
\nThe Gene Ontology organizes protein function into three axes: Molecular Function (what the protein does chemically), Biological Process (what larger goal it contributes to), and Cellular Component (where in the cell it operates). The same sequence can perform the same chemical activity in a completely different location while contributing to an entirely different process i.e the label is not a stable property of the sequence alone. On top of this, the output space is not a flat set of classes but a graph with tens of thousands of nodes, most proteins have very few labels, and crucially, a missing label does not mean \"no\" it means \"untested.\" You are being asked to learn a mapping where the inputs are the same across contexts, the outputs depend on context you don't have, and the training labels are incomplete by construction.
\nThere are a few terms about protiens I encourage you to look up so you understand why this is going to be really really hard. Mutlifunctionality aka \"moonlighting\", protiens serve multiple functions on many levels. For one thing the same amino acid sequence that allows it to perform a chemical reaction may actually be able to perform a slightly similar chemical rection and because biology is messy, it will happen. This means the sequence will not have a 1 to 1 MF mapping. Further more the molecular provcess itself appears in many different biological processes BP, and the protien in question will do this same MF for many differnt BPs as well as potentially a few different MFs in the same BP in siffernt contexts. Lastly you'd think the cellular component CC, where this is happening would be stable, but again its not. Protiens get shuffled around and end up in different locations.
\nI do not want to make it seem like there is no stability or predictability at all, its just that the really interesting biology does not happen here. The shared genes I mentioned earlier, they are the ones that are common across all biological contexts because when those get disrupted, the cell dies. Evolution works by repurposing what exists and this is true in disease states as well, like cancer.
\nMaybe some one here more intelligent than me can take this into account some how, just wanted to share my thoughts and point out its not simply a matter of applying the fanciest model here. You'll need to apply some domain expertise here as well.
", "rawMarkdown": "I just wanted to share a brief post here for the computer scientists about the task at hand who may lack an understanding of what is being asked of them. To start with we need some defintions of things you may think are farily obvious but are actually not. For starters what is a gene, what is a protien, and what are their functions? These are critical since the predictive task is to map amino acid sequence of a protien to its \"funciton\" i.e Gene Ontology terms. We need to get into the weeds here. The description in the data seciton, https://www.kaggle.com/competitions/cafa-6-protein-function-prediction/data left several important things out we need to go over.\n\nBasically you have DNA, composed of the 4 nucleotide bases which have a triplet code that gets read out into mRNA then into protien amino acid sequences. Here we use the amino acide sequences of the protien as input. Simple enough right? Except we need to consider where this data is coming from. Thousands of different experiments in differnt biological contexts, many different cell types in many different disease states in many different species and cell lines. The protiens are not unique to these contexts nor are the sequences. \n\nWhat I want you to consider is that it is well known that humans share 99% of our coding protien genes with chimpanzes our closest genetic relative. What is less widely known is that we also share ~80% with a bananna or for that matter nearly any complex mutlicellular life and a good deal of the single cellular life realted to mulicellular life, on the eukaryotic branch of our family tree. Once nature evolves a gene it does not like to remake that gene. Instead it gets repurposed, and there are different levels to what we as a human might term a \"purpose\" or \"function\". This brings us to the gene ontology.\n\nThe Gene Ontology organizes protein function into three axes: Molecular Function (what the protein does chemically), Biological Process (what larger goal it contributes to), and Cellular Component (where in the cell it operates). The same sequence can perform the same chemical activity in a completely different location while contributing to an entirely different process i.e the label is not a stable property of the sequence alone. On top of this, the output space is not a flat set of classes but a graph with tens of thousands of nodes, most proteins have very few labels, and crucially, a missing label does not mean \"no\" it means \"untested.\" You are being asked to learn a mapping where the inputs are the same across contexts, the outputs depend on context you don't have, and the training labels are incomplete by construction.\n\nThere are a few terms about protiens I encourage you to look up so you understand why this is going to be really really hard. Mutlifunctionality aka \"moonlighting\", protiens serve multiple functions on many levels. For one thing the same amino acid sequence that allows it to perform a chemical reaction may actually be able to perform a slightly similar chemical rection and because biology is messy, it will happen. This means the sequence will not have a 1 to 1 MF mapping. Further more the molecular provcess itself appears in many different biological processes BP, and the protien in question will do this same MF for many differnt BPs as well as potentially a few different MFs in the same BP in siffernt contexts. Lastly you'd think the cellular component CC, where this is happening would be stable, but again its not. Protiens get shuffled around and end up in different locations. \n\nI do not want to make it seem like there is no stability or predictability at all, its just that the really interesting biology does not happen here. The shared genes I mentioned earlier, they are the ones that are common across all biological contexts because when those get disrupted, the cell dies. Evolution works by repurposing what exists and this is true in disease states as well, like cancer. \n\nMaybe some one here more intelligent than me can take this into account some how, just wanted to share my thoughts and point out its not simply a matter of applying the fanciest model here. You'll need to apply some domain expertise here as well.\n\n", "votes": 2 } ], "comments": [], "raw_markdown_by_id": { "3419870": "I just wanted to share a brief post here for the computer scientists about the task at hand who may lack an understanding of what is being asked of them. To start with we need some defintions of things you may think are farily obvious but are actually not. For starters what is a gene, what is a protien, and what are their functions? These are critical since the predictive task is to map amino acid sequence of a protien to its \"funciton\" i.e Gene Ontology terms. We need to get into the weeds here. The description in the data seciton, https://www.kaggle.com/competitions/cafa-6-protein-function-prediction/data left several important things out we need to go over.\n\nBasically you have DNA, composed of the 4 nucleotide bases which have a triplet code that gets read out into mRNA then into protien amino acid sequences. Here we use the amino acide sequences of the protien as input. Simple enough right? Except we need to consider where this data is coming from. Thousands of different experiments in differnt biological contexts, many different cell types in many different disease states in many different species and cell lines. The protiens are not unique to these contexts nor are the sequences. \n\nWhat I want you to consider is that it is well known that humans share 99% of our coding protien genes with chimpanzes our closest genetic relative. What is less widely known is that we also share ~80% with a bananna or for that matter nearly any complex mutlicellular life and a good deal of the single cellular life realted to mulicellular life, on the eukaryotic branch of our family tree. Once nature evolves a gene it does not like to remake that gene. Instead it gets repurposed, and there are different levels to what we as a human might term a \"purpose\" or \"function\". This brings us to the gene ontology.\n\nThe Gene Ontology organizes protein function into three axes: Molecular Function (what the protein does chemically), Biological Process (what larger goal it contributes to), and Cellular Component (where in the cell it operates). The same sequence can perform the same chemical activity in a completely different location while contributing to an entirely different process i.e the label is not a stable property of the sequence alone. On top of this, the output space is not a flat set of classes but a graph with tens of thousands of nodes, most proteins have very few labels, and crucially, a missing label does not mean \"no\" it means \"untested.\" You are being asked to learn a mapping where the inputs are the same across contexts, the outputs depend on context you don't have, and the training labels are incomplete by construction.\n\nThere are a few terms about protiens I encourage you to look up so you understand why this is going to be really really hard. Mutlifunctionality aka \"moonlighting\", protiens serve multiple functions on many levels. For one thing the same amino acid sequence that allows it to perform a chemical reaction may actually be able to perform a slightly similar chemical rection and because biology is messy, it will happen. This means the sequence will not have a 1 to 1 MF mapping. Further more the molecular provcess itself appears in many different biological processes BP, and the protien in question will do this same MF for many differnt BPs as well as potentially a few different MFs in the same BP in siffernt contexts. Lastly you'd think the cellular component CC, where this is happening would be stable, but again its not. Protiens get shuffled around and end up in different locations. \n\nI do not want to make it seem like there is no stability or predictability at all, its just that the really interesting biology does not happen here. The shared genes I mentioned earlier, they are the ones that are common across all biological contexts because when those get disrupted, the cell dies. Evolution works by repurposing what exists and this is true in disease states as well, like cancer. \n\nMaybe some one here more intelligent than me can take this into account some how, just wanted to share my thoughts and point out its not simply a matter of applying the fanciest model here. You'll need to apply some domain expertise here as well.\n\n" } }