There still seems to be some confusion about the meaning of different color channels, and how they help us identify protein localization when only the green channel is our protein of interest. This writing is inspired by this question, but I've read several others that have been similar.
\n\nProtein location in these images are not expressed in local terms, meaning that we do not need to determine the GPS coordinates of individual molecules. Instead, we are determining from images their global localization as a quality that describes their distribution. There are tens of thousands of different proteins in each cell, and for many of them tens of thousands of copies. So when we say that a protein A is in nucleus, that doesn't mean necessarily that it is in a particular position within the nucleus (though it can be). Usually that means that many copies of that protein are found somewhere - and often everywhere - in the nucleus. Because of it, we can have two completely different proteins A and B, completely unrelated to one another, that are uniformly dispersed in the nucleus and give identical images under the fluorescent microscope. When we classify the localizations of proteins, it matters less to us what protein is being imaged, but rather the pattern of its distribution in cells. That's why a classifier built on protein A can later be used to determine a localization of a completely unrelated protein Z, so long as they have similar cellular distribution. And we don't need to know anything else about either protein.
\n\nThree extra colors serve as general landmarks that help us narrow down the localization. Think of them as land, air and ocean. Knowing that a person is on land helps us eliminate air and ocean, but still leaves lots of possibilities as to which continent and which country is that person's residence. The blue channel tells us where nuclei are. If our protein of interest (green channel) overlaps with blue color only, that narrows its localization to 6 choices that are all within nucleus:
\n\n0. Nucleoplasm\n1. Nuclear membrane\n2. Nucleoli\n3. Nucleoli fibrillar center\n4. Nuclear speckles\n5. Nuclear bodies\nIf our green channel signal overlaps ONLY with blue channel signal and nothing else, that eliminates 22 other possibilities that come after #5. If, on the other hand, our green signal has zero overlap with blue, that eliminates the 6 localizations listed above, but leaves open the others.
\n\nIn general, all localizations after #5 are SOMEWHERE in the cytosol. Now, they can be EVERYWHERE in the cytosol, in which case they will belong to #25. Or they can be in a discreet localization within the cytosol, which is one of many numbers after #5. That's where red and yellow channels come into play. The red channel shows cable-like proteins called microtubules, which represent cellular scaffolding. They are found only in cytosol, so they do not overlap with nucleus. If blue signal (nucleus) was land, red signal (microtubules) is the oceans. Microtubules are good cytoplasmic markers because they are uniformly distributed in all cells. The yellow channel is endoplasmatic reticulum, which is also within the cytosol, but has less uniform distribution than microtubules and doesn't necessarily cover every inch of the cytosol.
\n\nSo the pattern that has to be learned can be boiled down to the following: 1) are the proteins somewhere in nucleus or somewhere in cytosol? They can certainly be in both, but it helps if they are only in one of them. 2) if in nucleus, are they at the edge of blue signal (that would be nuclear membrane); if not, are they in some kind of globs/speckles? if yes, larger globs that number in 5 or fewer are usually nucleoli or nucleoli fibrillar centers. 3) if they are outside of nucleus, some are in organelles and others are in various protein complexes; if the former, mitochondria, endosomes and lysosomes tend to be uniformly distributed in cytosol but not fill it up completely; Golgi tends to be only at one side of the cell and usually close to nucleus, same for centrosome.
\n\nI am not covering all the possibilities here - that's something that takes more studying. I suggest you take a tour through various cellular localization in this interactive page as that should help understand the expected signal pattern for each particular localizations.
", "messages": [ { "id": 428960, "postDate": "2018-11-28T06:25:27.673Z", "content": "There still seems to be some confusion about the meaning of different color channels, and how they help us identify protein localization when only the green channel is our protein of interest. This writing is inspired by this question, but I've read several others that have been similar.
\n\nProtein location in these images are not expressed in local terms, meaning that we do not need to determine the GPS coordinates of individual molecules. Instead, we are determining from images their global localization as a quality that describes their distribution. There are tens of thousands of different proteins in each cell, and for many of them tens of thousands of copies. So when we say that a protein A is in nucleus, that doesn't mean necessarily that it is in a particular position within the nucleus (though it can be). Usually that means that many copies of that protein are found somewhere - and often everywhere - in the nucleus. Because of it, we can have two completely different proteins A and B, completely unrelated to one another, that are uniformly dispersed in the nucleus and give identical images under the fluorescent microscope. When we classify the localizations of proteins, it matters less to us what protein is being imaged, but rather the pattern of its distribution in cells. That's why a classifier built on protein A can later be used to determine a localization of a completely unrelated protein Z, so long as they have similar cellular distribution. And we don't need to know anything else about either protein.
\n\nThree extra colors serve as general landmarks that help us narrow down the localization. Think of them as land, air and ocean. Knowing that a person is on land helps us eliminate air and ocean, but still leaves lots of possibilities as to which continent and which country is that person's residence. The blue channel tells us where nuclei are. If our protein of interest (green channel) overlaps with blue color only, that narrows its localization to 6 choices that are all within nucleus:
\n\n0. Nucleoplasm\n1. Nuclear membrane\n2. Nucleoli\n3. Nucleoli fibrillar center\n4. Nuclear speckles\n5. Nuclear bodies\nIf our green channel signal overlaps ONLY with blue channel signal and nothing else, that eliminates 22 other possibilities that come after #5. If, on the other hand, our green signal has zero overlap with blue, that eliminates the 6 localizations listed above, but leaves open the others.
\n\nIn general, all localizations after #5 are SOMEWHERE in the cytosol. Now, they can be EVERYWHERE in the cytosol, in which case they will belong to #25. Or they can be in a discreet localization within the cytosol, which is one of many numbers after #5. That's where red and yellow channels come into play. The red channel shows cable-like proteins called microtubules, which represent cellular scaffolding. They are found only in cytosol, so they do not overlap with nucleus. If blue signal (nucleus) was land, red signal (microtubules) is the oceans. Microtubules are good cytoplasmic markers because they are uniformly distributed in all cells. The yellow channel is endoplasmatic reticulum, which is also within the cytosol, but has less uniform distribution than microtubules and doesn't necessarily cover every inch of the cytosol.
\n\nSo the pattern that has to be learned can be boiled down to the following: 1) are the proteins somewhere in nucleus or somewhere in cytosol? They can certainly be in both, but it helps if they are only in one of them. 2) if in nucleus, are they at the edge of blue signal (that would be nuclear membrane); if not, are they in some kind of globs/speckles? if yes, larger globs that number in 5 or fewer are usually nucleoli or nucleoli fibrillar centers. 3) if they are outside of nucleus, some are in organelles and others are in various protein complexes; if the former, mitochondria, endosomes and lysosomes tend to be uniformly distributed in cytosol but not fill it up completely; Golgi tends to be only at one side of the cell and usually close to nucleus, same for centrosome.
\n\nI am not covering all the possibilities here - that's something that takes more studying. I suggest you take a tour through various cellular localization in this interactive page as that should help understand the expected signal pattern for each particular localizations.
", "rawMarkdown": "There still seems to be some confusion about the meaning of different color channels, and how they help us identify protein localization when only the green channel is our protein of interest. This writing is inspired by [__this question__](https://www.kaggle.com/c/human-protein-atlas-image-classification/discussion/72673#428876), but I've read several others that have been similar.\n\nProtein location in these images are not expressed in local terms, meaning that we do not need to determine the GPS coordinates of individual molecules. Instead, we are determining from images their global localization as a quality that describes their distribution. There are tens of thousands of different proteins in each cell, and for many of them tens of thousands of copies. So when we say that a protein A is in nucleus, that doesn't mean necessarily that it is in a particular position within the nucleus (though it can be). Usually that means that many copies of that protein are found somewhere - and often everywhere - in the nucleus. Because of it, we can have two completely different proteins A and B, completely unrelated to one another, that are uniformly dispersed in the nucleus and give identical images under the fluorescent microscope. When we classify the localizations of proteins, it matters less to us what protein is being imaged, but rather the pattern of its distribution in cells. That's why a classifier built on protein A can later be used to determine a localization of a completely unrelated protein Z, so long as they have similar cellular distribution. And we don't need to know anything else about either protein.\n\nThree extra colors serve as general landmarks that help us narrow down the localization. Think of them as land, air and ocean. Knowing that a person is on land helps us eliminate air and ocean, but still leaves lots of possibilities as to which continent and which country is that person's residence. The blue channel tells us where nuclei are. If our protein of interest (green channel) overlaps with blue color only, that narrows its localization to 6 choices that are all within nucleus:\n\n 0. Nucleoplasm\n 1. Nuclear membrane\n 2. Nucleoli\n 3. Nucleoli fibrillar center\n 4. Nuclear speckles\n 5. Nuclear bodies\n\nIf our green channel signal overlaps ONLY with blue channel signal and nothing else, that eliminates 22 other possibilities that come after #5. If, on the other hand, our green signal has zero overlap with blue, that eliminates the 6 localizations listed above, but leaves open the others. \n\nIn general, all localizations after #5 are SOMEWHERE in the cytosol. Now, they can be EVERYWHERE in the cytosol, in which case they will belong to #25. Or they can be in a discreet localization within the cytosol, which is one of many numbers after #5. That's where red and yellow channels come into play. The red channel shows cable-like proteins called microtubules, which represent cellular scaffolding. They are found only in cytosol, so they do not overlap with nucleus. If blue signal (nucleus) was land, red signal (microtubules) is the oceans. Microtubules are good cytoplasmic markers because they are uniformly distributed in all cells. The yellow channel is endoplasmatic reticulum, which is also within the cytosol, but has less uniform distribution than microtubules and doesn't necessarily cover every inch of the cytosol.\n\nSo the pattern that has to be learned can be boiled down to the following: 1) are the proteins somewhere in nucleus or somewhere in cytosol? They can certainly be in both, but it helps if they are only in one of them. 2) if in nucleus, are they at the edge of blue signal (that would be nuclear membrane); if not, are they in some kind of globs/speckles? if yes, larger globs that number in 5 or fewer are usually nucleoli or nucleoli fibrillar centers. 3) if they are outside of nucleus, some are in organelles and others are in various protein complexes; if the former, mitochondria, endosomes and lysosomes tend to be uniformly distributed in cytosol but not fill it up completely; Golgi tends to be only at one side of the cell and usually close to nucleus, same for centrosome.\n\nI am not covering all the possibilities here - that's something that takes more studying. I suggest you take a tour through various cellular localization in this [__interactive page__](https://www.proteinatlas.org/learn/dictionary/cell) as that should help understand the expected signal pattern for each particular localizations.", "votes": 41 }, { "id": 432368, "postDate": "2018-12-03T19:23:08.847Z", "content": "Great commentary and insights. I have been looking at my model results on the test data (best LB score is .459) a lot and I am seeing some combinations of proteins that are never present in the training data so I am very suspicious of them. I have been training with green, red, and blue channels only.
\n\ntx
\n\nKickback
", "rawMarkdown": "Great commentary and insights. I have been looking at my model results on the test data (best LB score is .459) a lot and I am seeing some combinations of proteins that are never present in the training data so I am very suspicious of them. I have been training with green, red, and blue channels only.\n\ntx\n\nKickback", "votes": 1 }, { "id": 431955, "postDate": "2018-12-03T06:47:34.980Z", "content": "Really helpful
", "rawMarkdown": "Really helpful", "votes": 1 }, { "id": 430504, "postDate": "2018-11-30T13:36:58.970Z", "content": "I just wrote something to suppress impossible categories by setting them to 0 post-prediction. Currently just nuclear or cytoplasmic, as I don't want to be too restrictive. We are lucky that they are single plane confocal image slices, so in theory there shouldn't be cytoplasmic staining bleed-through in the nuclear channel. There would be cytoplasm behind / in front of the nucleus in a total luminescence image. i.e. cells are 3D entities.
\n\nMy model is taking forever to train, but I will report on whether it makes any difference to LB when I make my submits. I think the shifted channels you found may mean in some cases this might not help, but hopefully will be useful overall.
", "rawMarkdown": "I just wrote something to suppress impossible categories by setting them to 0 post-prediction. Currently just nuclear or cytoplasmic, as I don't want to be too restrictive. We are lucky that they are single plane confocal image slices, so *in theory* there shouldn't be cytoplasmic staining bleed-through in the nuclear channel. There would be cytoplasm behind / in front of the nucleus in a total luminescence image. i.e. cells are 3D entities.\n\nMy model is taking forever to train, but I will report on whether it makes any difference to LB when I make my submits. I think the shifted channels you found may mean in some cases this might not help, but hopefully will be useful overall.", "votes": 1, "replies": [ { "id": 430550, "postDate": "2018-11-30T15:17:31.410Z", "content": "\n\n\nI just wrote something to suppress impossible categories by setting them to 0 post-prediction.
\n
@dstjhb This is a good idea in general when one is certain that some combinations are infeasible. However, I have already seen some combinations that I consider unrealistic, and that's after inspecting only a small number of images. I don't want to go through them because in this case it doesn't matter what I think - we have to train based on the provided ground truth.
", "rawMarkdown": "> I just wrote something to suppress impossible categories by setting them to 0 post-prediction.\n\n@dstjhb This is a good idea in general when one is certain that some combinations are infeasible. However, I have already seen some combinations that I consider unrealistic, and that's after inspecting only a small number of images. I don't want to go through them because in this case it doesn't matter what I think - we have to train based on the provided ground truth.", "votes": 1 }, { "id": 440259, "postDate": "2018-12-17T09:50:10.327Z", "content": "Thanks for nice article..\nLooking at the composite image how can depict the protein ,what are the visual traits if there are any based on which we say here is the protein in Image. \nIs any highlighted flueroscent will say that is protein or any black dots inside the nucleus says this is protein
", "rawMarkdown": "Thanks for nice article..\nLooking at the composite image how can depict the protein ,what are the visual traits if there are any based on which we say here is the protein in Image. \nIs any highlighted flueroscent will say that is protein or any black dots inside the nucleus says this is protein", "votes": 1 } ] }, { "id": 429061, "postDate": "2018-11-28T09:41:02.877Z", "content": "Thanks for the explanation using a easy to understand analogy (land, sea, air)! \nYour explanation also helps me to understand why these different channels were chosen.
", "rawMarkdown": "Thanks for the explanation using a easy to understand analogy (land, sea, air)! \nYour explanation also helps me to understand why these different channels were chosen.", "votes": 1 }, { "id": 428975, "postDate": "2018-11-28T07:00:28.657Z", "content": "Oh wow i was so wrong in understanding what I was looking for... Luckily the neural networks still picked up the patterns correctly. Does it mean they are smarter than me?
", "rawMarkdown": "Oh wow i was so wrong in understanding what I was looking for... Luckily the neural networks still picked up the patterns correctly. Does it mean they are smarter than me? ", "votes": 1, "replies": [ { "id": 429007, "postDate": "2018-11-28T07:54:38.240Z", "content": "\n\n\nDoes it mean they are smarter than me?
\n
They are better than any of us at identifying patterns in a large volume of data. We are still a lot better at understanding the meaning of those patterns.
", "rawMarkdown": "> Does it mean they are smarter than me?\n\nThey are better than any of us at identifying patterns in a large volume of data. We are still a lot better at understanding the meaning of those patterns.", "votes": 3 } ] }, { "id": 429505, "postDate": "2018-11-29T00:25:41.797Z", "content": "Thanks for sharing. I am a molecular biologist and when I do my own inspection on the pictures I use similar approach in my head to identify the localization(s) of the protein of interest.
\n\nI think that the red and yellow channel are somehow pretty much the same when using this \"patterns of overlapping\" approach to find and calculate the features. But somehow in RGB mode red would be much easier to work with.
\n\nAlso we have noticed that some pictures in the training set are labelled slightly different from how they really look like they should get labelled. Some features, for example, the Xsome locations like #7, #8 and #9 are very similar to human eyes, biologist have some other approach to precisely map them but with only these three marker channel (blue, red and yellow), it would be really hard to tell them apart.
", "rawMarkdown": "Thanks for sharing. I am a molecular biologist and when I do my own inspection on the pictures I use similar approach in my head to identify the localization(s) of the protein of interest.\n\nI think that the red and yellow channel are somehow pretty much the same when using this \"patterns of overlapping\" approach to find and calculate the features. But somehow in RGB mode red would be much easier to work with.\n\nAlso we have noticed that some pictures in the training set are labelled slightly different from how they really look like they should get labelled. Some features, for example, the Xsome locations like #7, #8 and #9 are very similar to human eyes, biologist have some other approach to precisely map them but with only these three marker channel (blue, red and yellow), it would be really hard to tell them apart. ", "votes": 2 }, { "id": 428976, "postDate": "2018-11-28T07:01:36.370Z", "content": "Also, i have noticed that there are actually no photos with the relevant proteins marked. The interactive page is great.\nThanks again!
", "rawMarkdown": "Also, i have noticed that there are actually no photos with the relevant proteins marked. The interactive page is great.\nThanks again! " }, { "id": 431600, "postDate": "2018-12-02T15:13:37.517Z", "rawMarkdown": "", "votes": 2, "isDeleted": true }, { "id": 442324, "postDate": "2018-12-19T19:53:26.397Z", "content": "Thanks for sharing. Insightful.
", "rawMarkdown": "Thanks for sharing. Insightful.", "votes": 1 }, { "id": 431686, "postDate": "2018-12-02T18:36:54.790Z", "content": "Great information, Thank you!!
", "rawMarkdown": "Great information, Thank you!!", "votes": 1 }, { "id": 428987, "postDate": "2018-11-28T07:18:09.953Z", "content": "really helpful information, thank you.
", "rawMarkdown": "really helpful information, thank you.", "votes": 1 } ], "comments": [ { "id": 432368, "author_name": "kickback", "author_url": "", "post_date": "2018-12-03T19:23:08.847000", "content": "Great commentary and insights. I have been looking at my model results on the test data (best LB score is .459) a lot and I am seeing some combinations of proteins that are never present in the training data so I am very suspicious of them. I have been training with green, red, and blue channels only.
\n\ntx
\n\nKickback
", "votes": 1, "replies": [] }, { "id": 431955, "author_name": "CYBruce", "author_url": "", "post_date": "2018-12-03T06:47:34.980000", "content": "Really helpful
", "votes": 1, "replies": [] }, { "id": 430504, "author_name": "DStjhb", "author_url": "", "post_date": "2018-11-30T13:36:58.970000", "content": "I just wrote something to suppress impossible categories by setting them to 0 post-prediction. Currently just nuclear or cytoplasmic, as I don't want to be too restrictive. We are lucky that they are single plane confocal image slices, so in theory there shouldn't be cytoplasmic staining bleed-through in the nuclear channel. There would be cytoplasm behind / in front of the nucleus in a total luminescence image. i.e. cells are 3D entities.
\n\nMy model is taking forever to train, but I will report on whether it makes any difference to LB when I make my submits. I think the shifted channels you found may mean in some cases this might not help, but hopefully will be useful overall.
", "votes": 1, "replies": [ { "id": 430550, "author_name": "Tilii", "author_url": "", "post_date": "2018-11-30T15:17:31.410000", "content": "\n\n\nI just wrote something to suppress impossible categories by setting them to 0 post-prediction.
\n
@dstjhb This is a good idea in general when one is certain that some combinations are infeasible. However, I have already seen some combinations that I consider unrealistic, and that's after inspecting only a small number of images. I don't want to go through them because in this case it doesn't matter what I think - we have to train based on the provided ground truth.
", "votes": 1, "replies": [] }, { "id": 440259, "author_name": "Jaideep", "author_url": "", "post_date": "2018-12-17T09:50:10.327000", "content": "Thanks for nice article..\nLooking at the composite image how can depict the protein ,what are the visual traits if there are any based on which we say here is the protein in Image. \nIs any highlighted flueroscent will say that is protein or any black dots inside the nucleus says this is protein
", "votes": 1, "replies": [] } ] }, { "id": 429061, "author_name": "FabSchreiber", "author_url": "", "post_date": "2018-11-28T09:41:02.877000", "content": "Thanks for the explanation using a easy to understand analogy (land, sea, air)! \nYour explanation also helps me to understand why these different channels were chosen.
", "votes": 1, "replies": [] }, { "id": 428975, "author_name": "Moshel", "author_url": "", "post_date": "2018-11-28T07:00:28.657000", "content": "Oh wow i was so wrong in understanding what I was looking for... Luckily the neural networks still picked up the patterns correctly. Does it mean they are smarter than me?
", "votes": 1, "replies": [ { "id": 429007, "author_name": "Tilii", "author_url": "", "post_date": "2018-11-28T07:54:38.240000", "content": "\n\n\nDoes it mean they are smarter than me?
\n
They are better than any of us at identifying patterns in a large volume of data. We are still a lot better at understanding the meaning of those patterns.
", "votes": 3, "replies": [] } ] }, { "id": 429505, "author_name": "Erhezi", "author_url": "", "post_date": "2018-11-29T00:25:41.797000", "content": "Thanks for sharing. I am a molecular biologist and when I do my own inspection on the pictures I use similar approach in my head to identify the localization(s) of the protein of interest.
\n\nI think that the red and yellow channel are somehow pretty much the same when using this \"patterns of overlapping\" approach to find and calculate the features. But somehow in RGB mode red would be much easier to work with.
\n\nAlso we have noticed that some pictures in the training set are labelled slightly different from how they really look like they should get labelled. Some features, for example, the Xsome locations like #7, #8 and #9 are very similar to human eyes, biologist have some other approach to precisely map them but with only these three marker channel (blue, red and yellow), it would be really hard to tell them apart.
", "votes": 2, "replies": [] }, { "id": 428976, "author_name": "Moshel", "author_url": "", "post_date": "2018-11-28T07:01:36.370000", "content": "Also, i have noticed that there are actually no photos with the relevant proteins marked. The interactive page is great.\nThanks again!
", "votes": 0, "replies": [] }, { "id": 431600, "author_name": "", "author_url": "", "post_date": "2018-12-02T15:13:37.517000", "content": "", "votes": 2, "replies": [] }, { "id": 442324, "author_name": "Rokoson", "author_url": "", "post_date": "2018-12-19T19:53:26.397000", "content": "Thanks for sharing. Insightful.
", "votes": 1, "replies": [] }, { "id": 431686, "author_name": "Yonas Samuel", "author_url": "", "post_date": "2018-12-02T18:36:54.790000", "content": "Great information, Thank you!!
", "votes": 1, "replies": [] }, { "id": 428987, "author_name": "He Tao", "author_url": "", "post_date": "2018-11-28T07:18:09.953000", "content": "really helpful information, thank you.
", "votes": 1, "replies": [] } ], "raw_markdown_by_id": { "428960": "There still seems to be some confusion about the meaning of different color channels, and how they help us identify protein localization when only the green channel is our protein of interest. This writing is inspired by [__this question__](https://www.kaggle.com/c/human-protein-atlas-image-classification/discussion/72673#428876), but I've read several others that have been similar.\n\nProtein location in these images are not expressed in local terms, meaning that we do not need to determine the GPS coordinates of individual molecules. Instead, we are determining from images their global localization as a quality that describes their distribution. There are tens of thousands of different proteins in each cell, and for many of them tens of thousands of copies. So when we say that a protein A is in nucleus, that doesn't mean necessarily that it is in a particular position within the nucleus (though it can be). Usually that means that many copies of that protein are found somewhere - and often everywhere - in the nucleus. Because of it, we can have two completely different proteins A and B, completely unrelated to one another, that are uniformly dispersed in the nucleus and give identical images under the fluorescent microscope. When we classify the localizations of proteins, it matters less to us what protein is being imaged, but rather the pattern of its distribution in cells. That's why a classifier built on protein A can later be used to determine a localization of a completely unrelated protein Z, so long as they have similar cellular distribution. And we don't need to know anything else about either protein.\n\nThree extra colors serve as general landmarks that help us narrow down the localization. Think of them as land, air and ocean. Knowing that a person is on land helps us eliminate air and ocean, but still leaves lots of possibilities as to which continent and which country is that person's residence. The blue channel tells us where nuclei are. If our protein of interest (green channel) overlaps with blue color only, that narrows its localization to 6 choices that are all within nucleus:\n\n 0. Nucleoplasm\n 1. Nuclear membrane\n 2. Nucleoli\n 3. Nucleoli fibrillar center\n 4. Nuclear speckles\n 5. Nuclear bodies\n\nIf our green channel signal overlaps ONLY with blue channel signal and nothing else, that eliminates 22 other possibilities that come after #5. If, on the other hand, our green signal has zero overlap with blue, that eliminates the 6 localizations listed above, but leaves open the others. \n\nIn general, all localizations after #5 are SOMEWHERE in the cytosol. Now, they can be EVERYWHERE in the cytosol, in which case they will belong to #25. Or they can be in a discreet localization within the cytosol, which is one of many numbers after #5. That's where red and yellow channels come into play. The red channel shows cable-like proteins called microtubules, which represent cellular scaffolding. They are found only in cytosol, so they do not overlap with nucleus. If blue signal (nucleus) was land, red signal (microtubules) is the oceans. Microtubules are good cytoplasmic markers because they are uniformly distributed in all cells. The yellow channel is endoplasmatic reticulum, which is also within the cytosol, but has less uniform distribution than microtubules and doesn't necessarily cover every inch of the cytosol.\n\nSo the pattern that has to be learned can be boiled down to the following: 1) are the proteins somewhere in nucleus or somewhere in cytosol? They can certainly be in both, but it helps if they are only in one of them. 2) if in nucleus, are they at the edge of blue signal (that would be nuclear membrane); if not, are they in some kind of globs/speckles? if yes, larger globs that number in 5 or fewer are usually nucleoli or nucleoli fibrillar centers. 3) if they are outside of nucleus, some are in organelles and others are in various protein complexes; if the former, mitochondria, endosomes and lysosomes tend to be uniformly distributed in cytosol but not fill it up completely; Golgi tends to be only at one side of the cell and usually close to nucleus, same for centrosome.\n\nI am not covering all the possibilities here - that's something that takes more studying. I suggest you take a tour through various cellular localization in this [__interactive page__](https://www.proteinatlas.org/learn/dictionary/cell) as that should help understand the expected signal pattern for each particular localizations.", "432368": "Great commentary and insights. I have been looking at my model results on the test data (best LB score is .459) a lot and I am seeing some combinations of proteins that are never present in the training data so I am very suspicious of them. I have been training with green, red, and blue channels only.\n\ntx\n\nKickback", "431955": "Really helpful", "430504": "I just wrote something to suppress impossible categories by setting them to 0 post-prediction. Currently just nuclear or cytoplasmic, as I don't want to be too restrictive. We are lucky that they are single plane confocal image slices, so *in theory* there shouldn't be cytoplasmic staining bleed-through in the nuclear channel. There would be cytoplasm behind / in front of the nucleus in a total luminescence image. i.e. cells are 3D entities.\n\nMy model is taking forever to train, but I will report on whether it makes any difference to LB when I make my submits. I think the shifted channels you found may mean in some cases this might not help, but hopefully will be useful overall.", "429061": "Thanks for the explanation using a easy to understand analogy (land, sea, air)! \nYour explanation also helps me to understand why these different channels were chosen.", "428975": "Oh wow i was so wrong in understanding what I was looking for... Luckily the neural networks still picked up the patterns correctly. Does it mean they are smarter than me? ", "429505": "Thanks for sharing. I am a molecular biologist and when I do my own inspection on the pictures I use similar approach in my head to identify the localization(s) of the protein of interest.\n\nI think that the red and yellow channel are somehow pretty much the same when using this \"patterns of overlapping\" approach to find and calculate the features. But somehow in RGB mode red would be much easier to work with.\n\nAlso we have noticed that some pictures in the training set are labelled slightly different from how they really look like they should get labelled. Some features, for example, the Xsome locations like #7, #8 and #9 are very similar to human eyes, biologist have some other approach to precisely map them but with only these three marker channel (blue, red and yellow), it would be really hard to tell them apart. ", "428976": "Also, i have noticed that there are actually no photos with the relevant proteins marked. The interactive page is great.\nThanks again! ", "431600": "", "442324": "Thanks for sharing. Insightful.", "431686": "Great information, Thank you!!", "428987": "really helpful information, thank you." } }