Classification of Thyroid Peroxidase (TPO) Inhibitors Using Transfer Learning with SMILES Embeddings | JoVE Visualize

Area of Science:

Toxicology and Pharmacology
Computational Chemistry
Bioinformatics

Background:

Thyroid hormones (THs) are vital for mammalian physiology, impacting all organs.
Thyroid peroxidase (TPO) is essential for TH biosynthesis, and its inhibition is a key event in thyroid disruption.
Current experimental and computational methods for assessing TPO activity are limited, necessitating advanced computational approaches.

Purpose of the Study:

To evaluate the efficacy of SMILES embeddings from a specialized language model (SLM) for quantitative structure-activity relationship (QSAR) modeling of TPO inhibitors.
To compare the performance of transfer learning using deep neural networks (DNNs) with traditional molecular descriptors for TPO inhibitor classification.
To advance in silico methodologies for identifying potential agrochemical TPO inhibitors.

Main Methods:

Utilized SMILES embeddings generated by a pretrained Bidirectional Encoder Representations from Transformers (BERT) based SLM.
Developed Random Forest (RF) models using both SMILES embeddings and traditional molecular descriptors.
Performed training, testing, and external validation of the developed QSAR models.
Compared model predictions with experimental TPO inhibition data from regulatory laboratories.

Main Results:

RF models using traditional descriptors showed similar performance on training and test sets but varied significantly in external validation sensitivity (0.788 vs 0.490).
SMILES embeddings demonstrated a more robust representation of structural information, potentially expanding the applicability domain of QSAR models.
Predictions from the models showed good agreement with experimental data from EFSA and EU-NETVAL, though some conflicting estimates arose from other data sources.

Conclusions:

Transfer learning with BERT-based SMILES embeddings offers a promising alternative to traditional molecular descriptors for QSAR modeling of TPO inhibitors.
This in silico approach advances the development of new approach methodologies (NAMs) for efficient screening of large chemical libraries.
Further research is recommended for specific compounds to resolve conflicting experimental data and refine predictive models.