\n
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F2262596%2Fb65def8a857ca441c70a81ae3fef9129%2FScreenshot%202023-12-01%20200101.png?generation=1701457468337451&alt=media\")
Have you tried removing CD8+ T-cells from training set - it helped to uplift score ?
\nWhat can be the reason for that ?
\nThere are several versions of pyboost with uplift 0.002 public LB from that.
\nHowever failed to find local confirmation for that phenomena.
Have you tried to duplicate several times in train some other sample groups (e.g. cell types) ?
\nThere are some pyboost publics which do these tricks and LB score uplifts, but again cannot find local support for that.
Some EDA observations:
\nClustermap shows that CD8+ T-cells somewhat different from the others. That supports removing CD8+
But dimensional reductions shows that CD8+ T-cells and T-regulatory cells are quite similar. That contradicts removing CD8+ , especially that top pyboost public adds T-regulatory with doubled multiplicity.
Clustermap ( https://www.kaggle.com/code/alexandervc/op2-eda-baseline-s?scriptVersionId=147818286&cellId=21 )
\n
Dimensional reduction by Antonina Dolgorukova:
\n![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F2262596%2F5493453fd4eb096b4716c7ab9cbedba1%2Fphoto_2023-12-01_20-06-22.jpg?generation=1701458423525125&alt=media\")
Dimensional reduction by Dmitry Ershov:
\n{'NK cells': 'red', 'T cells CD4+': 'black', 'T cells CD8+': 'blue', 'T regulatory cells': 'green', 'B cells': 'gold', 'Myeloid cells': 'magenta'}
![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F2262596%2F3ea4de7b559f5bbe000e57b89d2938a4%2Fphoto_2023-12-01_20-05-37.jpg?generation=1701458499894076&alt=media\")
Have you tried removing CD8+ T-cells from training set - it helped to uplift score ?
\nWhat can be the reason for that ?
\nThere are several versions of pyboost with uplift 0.002 public LB from that.
\nHowever failed to find local confirmation for that phenomena.
Have you tried to duplicate several times in train some other sample groups (e.g. cell types) ?
\nThere are some pyboost publics which do these tricks and LB score uplifts, but again cannot find local support for that.
Some EDA observations:
\nClustermap shows that CD8+ T-cells somewhat different from the others. That supports removing CD8+
But dimensional reductions shows that CD8+ T-cells and T-regulatory cells are quite similar. That contradicts removing CD8+ , especially that top pyboost public adds T-regulatory with doubled multiplicity.
Clustermap ( https://www.kaggle.com/code/alexandervc/op2-eda-baseline-s?scriptVersionId=147818286&cellId=21 )
\n
Dimensional reduction by Antonina Dolgorukova:
\n
Dimensional reduction by Dmitry Ershov:
\n{'NK cells': 'red', 'T cells CD4+': 'black', 'T cells CD8+': 'blue', 'T regulatory cells': 'green', 'B cells': 'gold', 'Myeloid cells': 'magenta'}
I used the lower limit of positive control as threshold and divided each expression pattern into two parts; I excluded CD8 because its positive pattern is different from the others. The score went up a bit.![\"\"](\"https://www.googleapis.com/download/storage/v1/b/kaggle-forum-message-attachments/o/inbox%2F2646279%2F95f221bee5bec8b3473fdfb3c64a75ec%2FOC2_alex.png?generation=1701470106516263&alt=media\")
I find that the linear system suggests DE from the NK and T4 may be most transferable to that of B Cell and Myeloid cells - https://www.kaggle.com/code/makio323/37-a-model-linear-algebra-0-768-private-0-582-p
", "rawMarkdown": "I find that the linear system suggests DE from the NK and T4 may be most transferable to that of B Cell and Myeloid cells - https://www.kaggle.com/code/makio323/37-a-model-linear-algebra-0-768-private-0-582-p", "votes": null }, { "id": "2546734", "postDate": "12/02/2023 18:55:31", "content": "Thanks for sharing !
\nWould you be so kind to comment on \" lower limit of positive control \" - it is minimal for EACH gene ? (minimum across all positive control samples ?)
Would you be so kind to give an example \"positive pattern is different from the others\" ?
\nThanks in advance !
", "rawMarkdown": "Thanks for sharing ! \nWould you be so kind to comment on \" lower limit of positive control \" - it is minimal for EACH gene ? (minimum across all positive control samples ?)\n\nWould you be so kind to give an example \"positive pattern is different from the others\" ?\n\nThanks in advance !", "votes": null }, { "id": "2546762", "postDate": "12/02/2023 19:57:39", "content": "Thank you for your comments and questions.
\nAt first I did not know what \"positive control\" meant.
\nI thought that \"positive control\" might refer to a threshold value that is considered \"positive\" if it meets or exceeds a certain threshold value.
\nThe lower limit is the minimum value for each gene.
\nExamples are CEP18770 andMLN2238.
I am not an expert in this field, so I apologize if my understanding is different.
", "rawMarkdown": "Thank you for your comments and questions.\n\nAt first I did not know what \"positive control\" meant.\n\nI thought that \"positive control\" might refer to a threshold value that is considered \"positive\" if it meets or exceeds a certain threshold value.\n\nThe lower limit is the minimum value for each gene.\nExamples are CEP18770 andMLN2238.\n\n[🧬🧬OP2_about positive control](https://www.kaggle.com/code/yoshifumimiya/op2-about-positive-control)\n\nI am not an expert in this field, so I apologize if my understanding is different.", "votes": null } ], "comments": [ { "id": 2545917, "author_name": "yoshifumimiya", "author_url": "", "post_date": "12/01/2023 22:36:57", "content": "I used the lower limit of positive control as threshold and divided each expression pattern into two parts; I excluded CD8 because its positive pattern is different from the others. The score went up a bit.
Thanks for sharing !
\nWould you be so kind to comment on \" lower limit of positive control \" - it is minimal for EACH gene ? (minimum across all positive control samples ?)
Would you be so kind to give an example \"positive pattern is different from the others\" ?
\nThanks in advance !
", "votes": null, "replies": [ { "id": 2546762, "author_name": "yoshifumimiya", "author_url": "", "post_date": "12/02/2023 19:57:39", "content": "Thank you for your comments and questions.
\nAt first I did not know what \"positive control\" meant.
\nI thought that \"positive control\" might refer to a threshold value that is considered \"positive\" if it meets or exceeds a certain threshold value.
\nThe lower limit is the minimum value for each gene.
\nExamples are CEP18770 andMLN2238.
I am not an expert in this field, so I apologize if my understanding is different.
", "votes": null, "replies": [] } ] } ] }, { "id": 2546629, "author_name": "makio323", "author_url": "", "post_date": "12/02/2023 16:19:38", "content": "I find that the linear system suggests DE from the NK and T4 may be most transferable to that of B Cell and Myeloid cells - https://www.kaggle.com/code/makio323/37-a-model-linear-algebra-0-768-private-0-582-p
", "votes": null, "replies": [] } ], "raw_markdown_by_id": { "2545830": "Have you tried removing CD8+ T-cells from training set - it helped to uplift score ? \nWhat can be the reason for that ? \nThere are several versions of pyboost with uplift 0.002 public LB from that.\nHowever failed to find local confirmation for that phenomena.\n\nHave you tried to duplicate several times in train some other sample groups (e.g. cell types) ? \nThere are some pyboost publics which do these tricks and LB score uplifts, but again cannot find local support for that. \n\nSome EDA observations: \n\n1. Clustermap shows that CD8+ T-cells somewhat different from the others. That supports removing CD8+\n\n2. But dimensional reductions shows that CD8+ T-cells and T-regulatory cells are quite similar. That contradicts removing CD8+ , especially that top pyboost public adds T-regulatory with doubled multiplicity.\n\nClustermap ( https://www.kaggle.com/code/alexandervc/op2-eda-baseline-s?scriptVersionId=147818286&cellId=21 )\n\n\nDimensional reduction by Antonina Dolgorukova:\n\n\nDimensional reduction by Dmitry Ershov:\n{'NK cells': 'red', 'T cells CD4+': 'black', 'T cells CD8+': 'blue', 'T regulatory cells': 'green', 'B cells': 'gold', 'Myeloid cells': 'magenta'}\n\n", "2545917": "I used the lower limit of positive control as threshold and divided each expression pattern into two parts; I excluded CD8 because its positive pattern is different from the others. The score went up a bit.", "2546629": "I find that the linear system suggests DE from the NK and T4 may be most transferable to that of B Cell and Myeloid cells - https://www.kaggle.com/code/makio323/37-a-model-linear-algebra-0-768-private-0-582-p", "2546734": "Thanks for sharing ! \nWould you be so kind to comment on \" lower limit of positive control \" - it is minimal for EACH gene ? (minimum across all positive control samples ?)\n\nWould you be so kind to give an example \"positive pattern is different from the others\" ?\n\nThanks in advance !", "2546762": "Thank you for your comments and questions.\n\nAt first I did not know what \"positive control\" meant.\n\nI thought that \"positive control\" might refer to a threshold value that is considered \"positive\" if it meets or exceeds a certain threshold value.\n\nThe lower limit is the minimum value for each gene.\nExamples are CEP18770 andMLN2238.\n\n[🧬🧬OP2_about positive control](https://www.kaggle.com/code/yoshifumimiya/op2-about-positive-control)\n\nI am not an expert in this field, so I apologize if my understanding is different." }, "source": "meta" }