Uploading my notebook and relevant files. I will try to write something up more fully when I get the time. Summary:
\nI've made two categories of models basic (only using SMILES given) and basic-plus (using SMILES plus Id number). Basic-plus (in my mind) has a fairly similar architecture to basic and should have performed similarly, if Id wasn't a significant factor. However, the best (and latest) basic model got a score of 0.08981 and the best (and the one on leaderboard #6) got a score of 0.11536--a fairly wide delta.
\nBasic-plus:
\nLeaderboard submission is basic-plus (notebook attached, relevant file https://drive.google.com/file/d/1rBCAxlrbaWF983gBckcPQW2BMtzel1jS/view?usp=sharing).
Basic
\nThe best basic model (0.08981) is a slight modification of the approach presented in the discussion post (https://www.kaggle.com/code/davidoloren/getting-score-0-11353-multi-representation-model/notebook) which ended with a private score of 0.07443.
… I think the situation is very much complicated by noise in the underlying, and I would further share that to my knowledge the basic models are fairly representative of high-quality molecular property predictors, and similar models have scored very highly on the therapeutic data commons dataset
", "comments": [ { "id": 2124962, "author_name": "Bernhard Rohde", "author_url": "", "post_date": "2023-02-01T10:21:25.407000", "content": "@davidoloren Looks like your basic-plus model was catching some of the plate structure. Have you tried using an explicitly guessed plate number (as in my solution https://www.kaggle.com/code/bernhardrohde/plate-row-column-model-with-clogp-for-solubility) as a descriptor?
\nimport re\n# Assuming 16 columns by 22 rows, which is a flipped format\nNCOLS = 16\nNROWS = 22\n\ndef ID2Coordinates(id) :\n # returns a tripple of plate, row, column indexes inferred from the EOS id\n intID = int(re.search(r'\\d+', id).group())\n if intID > 100000 : # top-up run\n plate = 280 + (intID-100001)//352\n row = (intID-100001)%352//NCOLS\n column = (intID-100001)%352%NCOLS\n else :\n plate = (intID-1)//352\n row = (intID-1)%352//NCOLS\n column = (intID-1)%352%NCOLS\n return plate, row, column\n", "votes": 1, "replies": [ { "id": 2128668, "author_name": "DHuang", "author_url": "", "post_date": "2023-02-03T23:16:15.103000", "content": "
Nope you were way ahead of me on that one, I thought your guess was particularly insightful, I had seen similar trends based on a moving average of proportion of 0/1/2, but hadn't thought to explicitly demarcate them by an explicitly guessed plate number. Interestingly, for visualization I'd settled on using a moving average window of 400, similar to the 352, but I hadn't made to final leap (and I suspect that that actually is the main difference between the score you'd achieved and basic-plus).
", "votes": 2, "replies": [ { "id": 2128994, "author_name": "Bernhard Rohde", "author_url": "", "post_date": "2023-02-04T08:37:11.140000", "content": "I guess listening to the scientists doing the actual experiments for many years actually paid off here. Looks like experience (30+ years) can beat fancy methods at times ;-).
", "votes": 1, "replies": [] } ] } ] } ], "comment_count": 3, "messages": [ { "id": 2124029, "postDate": "2023-01-31T19:54:38.453Z", "content": "Uploading my notebook and relevant files. I will try to write something up more fully when I get the time. Summary:
\nI've made two categories of models basic (only using SMILES given) and basic-plus (using SMILES plus Id number). Basic-plus (in my mind) has a fairly similar architecture to basic and should have performed similarly, if Id wasn't a significant factor. However, the best (and latest) basic model got a score of 0.08981 and the best (and the one on leaderboard #6) got a score of 0.11536--a fairly wide delta.
\nBasic-plus:
\nLeaderboard submission is basic-plus (notebook attached, relevant file https://drive.google.com/file/d/1rBCAxlrbaWF983gBckcPQW2BMtzel1jS/view?usp=sharing).
Basic
\nThe best basic model (0.08981) is a slight modification of the approach presented in the discussion post (https://www.kaggle.com/code/davidoloren/getting-score-0-11353-multi-representation-model/notebook) which ended with a private score of 0.07443.
… I think the situation is very much complicated by noise in the underlying, and I would further share that to my knowledge the basic models are fairly representative of high-quality molecular property predictors, and similar models have scored very highly on the therapeutic data commons dataset
", "rawMarkdown": "Uploading my notebook and relevant files. I will try to write something up more fully when I get the time. Summary:\n\nI've made two categories of models basic (only using SMILES given) and basic-plus (using SMILES plus Id number). Basic-plus (in my mind) has a fairly similar architecture to basic and *should* have performed similarly, if Id wasn't a significant factor. However, the best (and latest) basic model got a score of 0.08981 and the best (and the one on leaderboard #6) got a score of 0.11536--a fairly wide delta.\n\n**Basic-plus:**\nLeaderboard submission is basic-plus (notebook attached, relevant file https://drive.google.com/file/d/1rBCAxlrbaWF983gBckcPQW2BMtzel1jS/view?usp=sharing).\n\n**Basic**\nThe best basic model (0.08981) is a slight modification of the approach presented in the discussion post (https://www.kaggle.com/code/davidoloren/getting-score-0-11353-multi-representation-model/notebook) which ended with a private score of 0.07443.\n\n... I think the situation is very much complicated by noise in the underlying, and I would further share that to my knowledge the basic models are fairly representative of high-quality molecular property predictors, and similar models have scored very highly on the therapeutic data commons dataset", "votes": 5 }, { "id": 2124962, "postDate": "2023-02-01T10:21:25.407Z", "content": "@davidoloren Looks like your basic-plus model was catching some of the plate structure. Have you tried using an explicitly guessed plate number (as in my solution https://www.kaggle.com/code/bernhardrohde/plate-row-column-model-with-clogp-for-solubility) as a descriptor?
\nimport re\n# Assuming 16 columns by 22 rows, which is a flipped format\nNCOLS = 16\nNROWS = 22\n\ndef ID2Coordinates(id) :\n # returns a tripple of plate, row, column indexes inferred from the EOS id\n intID = int(re.search(r'\\d+', id).group())\n if intID > 100000 : # top-up run\n plate = 280 + (intID-100001)//352\n row = (intID-100001)%352//NCOLS\n column = (intID-100001)%352%NCOLS\n else :\n plate = (intID-1)//352\n row = (intID-1)%352//NCOLS\n column = (intID-1)%352%NCOLS\n return plate, row, column\n", "rawMarkdown": "@davidoloren Looks like your basic-plus model was catching some of the plate structure. Have you tried using an explicitly guessed plate number (as in my solution https://www.kaggle.com/code/bernhardrohde/plate-row-column-model-with-clogp-for-solubility) as a descriptor?\n
\nimport re\n# Assuming 16 columns by 22 rows, which is a flipped format\nNCOLS = 16\nNROWS = 22\n\ndef ID2Coordinates(id) :\n # returns a tripple of plate, row, column indexes inferred from the EOS id\n intID = int(re.search(r'\\d+', id).group())\n if intID > 100000 : # top-up run\n plate = 280 + (intID-100001)//352\n row = (intID-100001)%352//NCOLS\n column = (intID-100001)%352%NCOLS\n else :\n plate = (intID-1)//352\n row = (intID-1)%352//NCOLS\n column = (intID-1)%352%NCOLS\n return plate, row, column\n", "votes": 1, "replies": [ { "id": 2128668, "postDate": "2023-02-03T23:16:15.103Z", "content": "
Nope you were way ahead of me on that one, I thought your guess was particularly insightful, I had seen similar trends based on a moving average of proportion of 0/1/2, but hadn't thought to explicitly demarcate them by an explicitly guessed plate number. Interestingly, for visualization I'd settled on using a moving average window of 400, similar to the 352, but I hadn't made to final leap (and I suspect that that actually is the main difference between the score you'd achieved and basic-plus).
", "rawMarkdown": "Nope you were way ahead of me on that one, I thought your guess was particularly insightful, I had seen similar trends based on a moving average of proportion of 0/1/2, but hadn't thought to explicitly demarcate them by an explicitly guessed plate number. Interestingly, for visualization I'd settled on using a moving average window of 400, similar to the 352, but I hadn't made to final leap (and I suspect that that actually is the main difference between the score you'd achieved and basic-plus).", "votes": 2, "replies": [ { "id": 2128994, "postDate": "2023-02-04T08:37:11.140Z", "content": "I guess listening to the scientists doing the actual experiments for many years actually paid off here. Looks like experience (30+ years) can beat fancy methods at times ;-).
", "rawMarkdown": "I guess listening to the scientists doing the actual experiments for many years actually paid off here. Looks like experience (30+ years) can beat fancy methods at times ;-).", "votes": 1 } ] } ] } ], "complete": true, "source": "live", "fetched_at": "2026-10-19T16:45:15Z", "competition_id": 38158, "competition_slug": "euos-slas", "team_id": 9187375, "solution_url": "/competitions/euos-slas/writeups/david-huang-draft-solution" }