Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Molecular Fingerprints Are Strong Models for Peptide Function Prediction

Jakub Adamczyk, Piotr Ludynia, Wojciech Czech

TL;DR

This work investigates whether long-range molecular dependencies are necessary for peptide function prediction. It introduces a simple, domain-specific encoding using count-based hashed molecular fingerprints (ECFP, Topological Torsion, RDKit) derived from sequence-based atom-level graphs, paired with LightGBM. Across 132 datasets and six benchmarks, these localized fingerprints achieve state-of-the-art performance, often outperforming long-range graph neural networks and pretrained protein language models while offering substantial speed and interpretability advantages. The findings suggest that short-range subgraph statistics capture the essential features for many peptide properties, providing an efficient and scalable alternative for peptide prediction with broad practical impact in drug discovery and peptide design.

Abstract

Understanding peptide properties is often assumed to require modeling long-range molecular interactions, motivating the use of complex graph neural networks and pretrained transformers. Yet, whether such long-range dependencies are essential remains unclear. We investigate if simple, domain-specific molecular fingerprints can capture peptide function without these assumptions. Atomic-level representation aims to provide richer information than purely sequence-based models and better efficiency than structural ones. Across 132 datasets, including LRGB and five other peptide benchmarks, models using count-based ECFP, Topological Torsion, and RDKit fingerprints with LightGBM achieve state-of-the-art accuracy. Despite encoding only short-range molecular features, these models outperform GNNs and transformer-based approaches. Control experiments with sequence shuffling and amino acid counts confirm that fingerprints, though inherently local, suffice for robust peptide property prediction. Our results challenge the presumed necessity of long-range interaction modeling and highlight molecular fingerprints as efficient, interpretable, and computationally lightweight alternatives for peptide prediction.

Molecular Fingerprints Are Strong Models for Peptide Function Prediction

TL;DR

This work investigates whether long-range molecular dependencies are necessary for peptide function prediction. It introduces a simple, domain-specific encoding using count-based hashed molecular fingerprints (ECFP, Topological Torsion, RDKit) derived from sequence-based atom-level graphs, paired with LightGBM. Across 132 datasets and six benchmarks, these localized fingerprints achieve state-of-the-art performance, often outperforming long-range graph neural networks and pretrained protein language models while offering substantial speed and interpretability advantages. The findings suggest that short-range subgraph statistics capture the essential features for many peptide properties, providing an efficient and scalable alternative for peptide prediction with broad practical impact in drug discovery and peptide design.

Abstract

Understanding peptide properties is often assumed to require modeling long-range molecular interactions, motivating the use of complex graph neural networks and pretrained transformers. Yet, whether such long-range dependencies are essential remains unclear. We investigate if simple, domain-specific molecular fingerprints can capture peptide function without these assumptions. Atomic-level representation aims to provide richer information than purely sequence-based models and better efficiency than structural ones. Across 132 datasets, including LRGB and five other peptide benchmarks, models using count-based ECFP, Topological Torsion, and RDKit fingerprints with LightGBM achieve state-of-the-art accuracy. Despite encoding only short-range molecular features, these models outperform GNNs and transformer-based approaches. Control experiments with sequence shuffling and amino acid counts confirm that fingerprints, though inherently local, suffice for robust peptide property prediction. Our results challenge the presumed necessity of long-range interaction modeling and highlight molecular fingerprints as efficient, interpretable, and computationally lightweight alternatives for peptide prediction.

Paper Structure

This paper contains 22 sections, 4 figures, 8 tables.

Table of Contents

  1. Introduction
  2. Related works
  3. Methods
  4. Molecular fingerprints
  5. Long-range interactions (LRI)
  6. Experiments and results
  7. Long Range Graph Benchmark (LRGB)
  8. Antimicrobial peptides (AMPs) benchmarks
  9. BERT-based models benchmark
  10. XUAMP benchmark
  11. AMPBenchmark
  12. General peptide benchmarks
  13. AutoPeptideML
  14. PeptideReactor
  15. Sequence shuffling analysis
  16. ...and 7 more sections

Figures (4)

  • Figure 1: The proposed pipeline for peptide function prediction.
  • Figure 2: Example subgraph extraction in consecutive iterations of ECFP algorithm. Reprinted with permission from ECFP. Copyright 2010 American Chemical Society.
  • Figure 3: Atom-level molecular graph for Liraglutide peptide.
  • Figure 4: Metrics for increasing shuffled amino acids ratio.