[Paper Review] Topological fingerprints reveal protein-ligand binding mechanism

Zixuan Cang, Guo‐Wei Wei|arXiv (Cornell University)|Mar 31, 2017

Topological and Geometric Data Analysis16 references4 citations

TL;DR

This paper introduces element-specific persistent homology (ESPH) and topological fingerprints (TFs) to capture essential geometric and chemical features of protein-ligand complexes, integrating them with machine learning for accurate binding affinity prediction. The proposed topology-based scoring function (T-Score) achieves a median Pearson correlation of 0.817 and RMSE of 1.912 kcal/mol on benchmark datasets, outperforming existing methods by effectively balancing topological abstraction with biological specificity.

ABSTRACT

Protein-ligand binding is a fundamental biological process that is paramount to many other biological processes, such as signal transduction, metabolic pathways, enzyme construction, cell secretion, gene expression, etc. Accurate prediction of protein-ligand binding affinities is vital to rational drug design and the understanding of protein-ligand binding and binding induced function. Existing binding affinity prediction methods are inundated with geometric detail and involve excessively high dimensions, which undermines their predictive power for massive binding data. Topology provides an ultimate level of abstraction and thus incurs too much reduction in geometric information. Persistent homology embeds geometric information into topological invariants and bridges the gap between complex geometry and abstract topology. However, it over simplifies biological information. This work introduces element specific persistent homology (ESPH) to retain crucial biological information during topological simplification. The combination of ESPH and machine learning gives rise to one of the most efficient and powerful tools for revealing protein-ligand binding mechanism and for predicting binding affinities.

Motivation & Objective

• To address the limitations of existing scoring functions in predicting protein-ligand binding affinities due to excessive geometric detail or loss of biological specificity.
• To bridge the gap between high-dimensional geometric models and abstract topological abstractions by embedding multiscale geometric information into topological invariants.
• To develop a low-dimensional, biologically meaningful representation of protein-ligand interactions that retains crucial chemical and structural information for accurate prediction.
• To reveal the molecular mechanisms underlying protein-ligand binding through topological analysis of interaction networks and spatial scales.
• To demonstrate that topological fingerprints derived from persistent homology significantly improve binding affinity prediction performance on large benchmark datasets.

Proposed method

• Element-specific persistent homology (ESPH) is used to compute topological invariants (Betti numbers) separately for different chemical elements, preserving element-specific biological information during topological abstraction.
• Correlation function-based filtrations, including Lorentz and exponential functions, are applied to model multiscale interactions, with parameters (ν, τ) tuning the effective interaction length scales.
• Binned barcode representation (BBR) is introduced to convert persistent homology barcodes into compact, machine-learning-ready features by binning Betti numbers across scale intervals.
• Interactive persistent homology (IPH) is used to describe protein-ligand interactions, particularly focusing on C–C networks and their topological persistence.
• The T-Score model combines Betti-0, Betti-1, and Betti-2 features from all heavy atoms, carbon-specific ESTFs, and multiscale correlation-based filtrations to generate a low-dimensional, predictive feature set.
• A 500-fold random cross-validation is performed to assess the robustness and generalization of the T-Score model on the PDBBind v2007 core set.

Experimental results

Research questions

• RQ1 How can topological invariants derived from persistent homology be adapted to retain biologically relevant chemical information in protein-ligand complexes?
• RQ2 What is the optimal balance between topological abstraction and geometric specificity for accurate binding affinity prediction?
• RQ3 Which interaction scales and residue layers (e.g., first or second shell) are most predictive of protein-ligand binding affinity?
• RQ4 To what extent do long-range C–C interactions (beyond 40 Å) contribute to binding affinity prediction?
• RQ5 Can topological fingerprints outperform existing physics-based, knowledge-based, and empirical scoring functions in large-scale binding affinity prediction?

Key findings

• The T-Score model achieves a median Pearson correlation coefficient of 0.817 and an RMSE of 1.912 kcal/mol on the PDBBind v2007 core set, significantly outperforming existing methods.
• Betti-0 ESTFs derived from Lorentz correlation functions with (ν, τ) = (3, 2) yield a Pearson correlation of 0.769, while the exponential function with κ = τ = 1 achieves 0.782.
• Combining three sets of Betti-0 ESTFs at (ν, τ) = (3, 0.5), (3, 1), and (3, 2) increases the correlation to 0.784, demonstrating the benefit of multiscale analysis.
• Oscillatory peaks in prediction accuracy at τ ≈ 1.6 and 3.1 indicate the critical role of the first and second coordination shells of residues in binding affinity.
• Protein-ligand C–C interactions in the first two layers contribute significantly to prediction, but their impact diminishes between 3–5 τ.
• At large length scales (τ > 5), protein-ligand C–C Betti-1 and Betti-2 features show a surprising resurgence in predictive power, indicating long-range hydrophobic effects extend beyond 40 Å from the binding site.