{ "id": 512368, "title": "Boost Your Sequence Models with AtomInSmiles (AIS) Tokenization", "url": "/competitions/leash-BELKA/discussion/512368", "author_name": "Frenio Redeker", "post_date": "2024-06-14T20:45:18.727000", "votes": 20, "comment_count": 0, "views": 0, "content": "

Dear BELKA Competitors,

\n

If you're using sequence models in this competition, you might be interested in the AtomInSmiles (AIS) tokenization scheme, proposed by Ucak et al. in their recent paper (https://doi.org/10.1186/s13321-023-00725-9).

\n

AIS is an expressive tokenization method that accounts for all neighboring atoms of an encoded atom in a given molecular structure leading to a more diverse vocabulary and improved performance compared to basic SMILES tokenization. In my experiments, AIS tokenization led to a 14-50 % increase in average precision score compared to basic SMILES tokenization.

\n

To help you get started with AIS quickly, I have created the following resources:

\n
    \n
  1. A Kaggle dataset containing preprocessed (numericalized and zero-padded) AIS tokenized data.
    2. \n
  2. A Kaggle notebook comparing AIS to basic SMILES tokenization.
    3. \n
\n

If you have any questions or would like to discuss further, feel free to leave a comment below.

\n

Best of luck in the final stages of the competition!

", "messages": [ { "id": 2872512, "postDate": "2024-06-14T20:45:18.727Z", "content": "

Dear BELKA Competitors,

\n

If you're using sequence models in this competition, you might be interested in the AtomInSmiles (AIS) tokenization scheme, proposed by Ucak et al. in their recent paper (https://doi.org/10.1186/s13321-023-00725-9).

\n

AIS is an expressive tokenization method that accounts for all neighboring atoms of an encoded atom in a given molecular structure leading to a more diverse vocabulary and improved performance compared to basic SMILES tokenization. In my experiments, AIS tokenization led to a 14-50 % increase in average precision score compared to basic SMILES tokenization.

\n

To help you get started with AIS quickly, I have created the following resources:

\n
    \n
  1. A Kaggle dataset containing preprocessed (numericalized and zero-padded) AIS tokenized data.
    2. \n
  2. A Kaggle notebook comparing AIS to basic SMILES tokenization.
    3. \n
\n

If you have any questions or would like to discuss further, feel free to leave a comment below.

\n

Best of luck in the final stages of the competition!

", "rawMarkdown": "Dear BELKA Competitors,\n\nIf you're using sequence models in this competition, you might be interested in the AtomInSmiles (AIS) tokenization scheme, proposed by Ucak et *al.* in their recent paper ([https://doi.org/10.1186/s13321-023-00725-9](https://doi.org/10.1186/s13321-023-00725-9)). \n\nAIS is an expressive tokenization method that accounts for all neighboring atoms of an encoded atom in a given molecular structure leading to a more diverse vocabulary and improved performance compared to basic SMILES tokenization. In [my experiments](https://www.kaggle.com/code/frenio/leashbio-belka-smiles-vs-ais), AIS tokenization led to a 14-50 % increase in [average precision](https://lightning.ai/docs/torchmetrics/stable/classification/average_precision.html) score compared to basic SMILES tokenization.\n\nTo help you get started with AIS quickly, I have created the following resources:\n\n1. [A Kaggle dataset containing preprocessed (numericalized and zero-padded) AIS tokenized data](https://www.kaggle.com/datasets/frenio/leashbio-belka-numericalized-smiles-and-ais).\n2. [A Kaggle notebook comparing AIS to basic SMILES tokenization](https://www.kaggle.com/code/frenio/leashbio-belka-smiles-vs-ais).\n\nIf you have any questions or would like to discuss further, feel free to leave a comment below.\n\nBest of luck in the final stages of the competition!", "votes": 20 } ], "comments": [], "raw_markdown_by_id": { "2872512": "Dear BELKA Competitors,\n\nIf you're using sequence models in this competition, you might be interested in the AtomInSmiles (AIS) tokenization scheme, proposed by Ucak et *al.* in their recent paper ([https://doi.org/10.1186/s13321-023-00725-9](https://doi.org/10.1186/s13321-023-00725-9)). \n\nAIS is an expressive tokenization method that accounts for all neighboring atoms of an encoded atom in a given molecular structure leading to a more diverse vocabulary and improved performance compared to basic SMILES tokenization. In [my experiments](https://www.kaggle.com/code/frenio/leashbio-belka-smiles-vs-ais), AIS tokenization led to a 14-50 % increase in [average precision](https://lightning.ai/docs/torchmetrics/stable/classification/average_precision.html) score compared to basic SMILES tokenization.\n\nTo help you get started with AIS quickly, I have created the following resources:\n\n1. [A Kaggle dataset containing preprocessed (numericalized and zero-padded) AIS tokenized data](https://www.kaggle.com/datasets/frenio/leashbio-belka-numericalized-smiles-and-ais).\n2. [A Kaggle notebook comparing AIS to basic SMILES tokenization](https://www.kaggle.com/code/frenio/leashbio-belka-smiles-vs-ais).\n\nIf you have any questions or would like to discuss further, feel free to leave a comment below.\n\nBest of luck in the final stages of the competition!" } }