加载中…
加载中…
MLM RoBERTa-based pretrained model. Ready to fine-tune on specific tasks.
MLM RoBERTa-based pretrained model. 2 million of Simplified molecular-input line-entry system (SMILES) were used and BPE as tokenizer.
The model instended use is for fine-tuning on dowstream tasks were SMILES is the main input.
The model can be directly used for the classification of chemical compounds and prediction of molecular properties using SMILES representations.
The model can be fine-tuned for specific tasks such as drug discovery, toxicity prediction, and other cheminformatics applications using specific datasets.
The model should not be used for tasks outside of cheminformatics or without proper validation for the specific task. Misuse includes using the model for generating invalid chemical compounds or predictions outside the domain of trained data. Only works with SMILES, for SELFIES search miekmayuare repository.
The model may inherit biases from the training data. Limitations include potential overfitting to the pre-training tasks and resource intensity for training and fine-tuning.
2 million SELFIES were used to pretrain the model in order to mitigate missrepresentation (over and under-representation) of any type of molecules. Validation on known datasets for downstream tasks is the best way to see its limitations.
Use the code below to get started with the model.
from transformers import AutoModel, AutoTokenizer
tokenizer = AutoTokenizer.from_pretrained("mikemayuare/SMILYBPE")
model = AutoModel.from_pretrained("mikemayuare/SMILYBPE")
The training data comprised 2 million molecules from the PubChem dataset.
The models were pre-trained for 20 epochs using the AdamW optimizer on an NVIDIA 3060 GPU with 12GiB of VRAM.
SMILES were canonicalized and tokenizer was trained on a subset of 1 million molecules from the PubChem dataset.
Training time was approximately 72 hours on the specified hardware. Checkpoint sizes are approximately 500MB each.
Testing was conducted on MoleculeNet datasets, specifically BBBP, HIV, and Tox21.
Evaluation metrics were disaggregated by dataset and task type (e.g., binary classification for BBBP).
The primary evaluation metric was the ROC-AUC score, which is commonly used for binary classification tasks in cheminformatics (on fine-tuned models).
The models tokenized with APE generally outperformed those tokenized with BPE. SMILES models showed better performance than SELFIES models in most cases.
The model achieved competitive performance on standard benchmarks, outperforming several baseline models in specific tasks.
Interpretability analyses showed that models tokenized with APE preserved the chemical context better than those with BPE, leading to higher classification accuracy.
Carbon emissions were estimated using the Machine Learning Impact calculator.
The model architecture is based on RoBERTa with 6 hidden layers, 768 hidden size, 1536 intermediate size, and 12 attention heads.
BibTeX:
@mastersthesis{leon2024chemical,
title={Chemical Language Modeling},
author={Miguelangel Augusto Leon Mayuare},
year={2024},
school={NOVA Information Management School}
}
APA:
Mayuare, M. A. L. (2024). Chemical Language Modeling (Master's thesis). NOVA Information Management School.
SELFIES: A string-based representation of molecules. SMILES: Simplified Molecular Input Line Entry System, a notation for describing the structure of chemical species.
For more details, refer to the (pending publication)
For inquiries, please contact migueleonm@gmail.com