Molecular fingerprints are strong models for peptide function prediction

Area of Science:

• Computational chemistry
• Bioinformatics
• Machine learning

Background:

• Predicting peptide properties often relies on complex models assuming long-range molecular interactions.
• The necessity of modeling these long-range dependencies for peptide function remains uncertain.
• Investigating simpler, domain-specific molecular fingerprints as an alternative is crucial.

Purpose of the Study:

• To evaluate if molecular fingerprints can predict peptide function without complex long-range interaction modeling.
• To compare the performance of fingerprint-based models against graph neural networks (GNNs) and transformers.
• To determine if local molecular features are sufficient for robust peptide property prediction.

Main Methods:

• Utilized count-based Extended Connectivity Fingerprints (ECFP), Topological Torsion, and RDKit fingerprints.
• Employed LightGBM machine learning algorithm for classification tasks.
• Tested models across 132 datasets, including LRGB and peptide benchmarks.

Main Results:

• Fingerprint-based models achieved state-of-the-art accuracy on peptide property prediction tasks.
• These models outperformed complex GNNs and transformer-based approaches.
• Control experiments validated that local molecular features encoded by fingerprints are sufficient.

Conclusions:

• Simple molecular fingerprints are effective and efficient for peptide property prediction.
• The necessity of modeling long-range interactions for peptide function is challenged.
• Molecular fingerprints offer a lightweight, interpretable, and accurate alternative to complex deep learning models.