Using molecular embeddings in QSAR modeling: does it make a difference?

Area of Science:

• Computational chemistry
• cheminformatics
• drug discovery

Background:

• Deep learning has spurred novel molecular representation algorithms for drug discovery.
• Comparing these molecular embeddings against traditional methods for Quantitative Structure-Activity Relationship (QSAR) modeling is challenging.
• A fair comparison requires extensive experiments across diverse datasets and training scenarios.

Purpose of the Study:

• To systematically compare the performance of recently proposed molecular embedding techniques against traditional molecular representations in QSAR modeling.
• To provide empirical evidence to guide the selection of molecular representations for drug design tasks.

Main Methods:

• Literature review and reproduction of five recent molecular embedding methods (three unsupervised, two supervised).
• Comparative analysis of embedding performance on various QSAR classification and regression datasets.
• Benchmarking against traditional molecular descriptors and fingerprints using over 25,000 trained models.

Main Results:

• Molecular embeddings did not significantly outperform traditional molecular representations in QSAR tasks.
• Supervised molecular embeddings showed competitive performance.
• Unsupervised molecular embeddings generally performed worse than traditional representations.

Conclusions:

• The predictive power of molecular embeddings in QSAR modeling is not consistently superior to traditional methods.
• Thorough evaluation and comparison of embedding techniques are crucial before their application in computer-aided drug design.
• Further research is needed to fully understand and leverage the potential of molecular embeddings.