Novel Descriptors and Digital Signal Processing- Based Method for Protein Sequence Activity Relationship Study

Nicolas Fontaine, Xavier F. Cadet, Iyanar VetrivelInternational Journal of Molecular Sciences 2019 · DOI

TL;DR

This paper introduces "innov'SAR," a novel method for predicting protein function from sequence alone, by creating "extended numerical sequences" (Ext_SEQ) that combine multiple physicochemical descriptors and Fast Fourier Transform (FFT) analysis. This approach significantly improves prediction accuracy for various protein activities, offering a valuable tool for protein engineering and directed evolution.

What Problem Does This Paper Address?

Predicting how a protein's amino acid sequence relates to its function (its "fitness") is a central challenge in protein engineering and drug discovery. While structural information can be helpful, it's often unavailable or costly to obtain. Existing sequence-based methods can be computationally intensive, especially when trying to capture complex interactions between amino acids. The authors aim to develop a more efficient and accurate method for quantitative structure-activity relationship (QSAR) studies that leverages readily available sequence data and advanced signal processing techniques.

Methods

The core of the innov'SAR method lies in creating "extended numerical sequences" (Ext_SEQ) from protein sequences. This involves several steps:

  1. Elementary Numerical Sequences (Ele_SEQ): Amino acid sequences are first encoded into numerical representations using various physicochemical properties from the AAindex database.
  2. Fast Fourier Transform (FFT): For some of these numerical sequences, an FFT is applied to generate a "protein spectrum." This spectrum captures frequency-domain information, which the authors hypothesize can reveal interactions between amino acids that are not apparent in simple linear sequences.
  3. Extended Numerical Sequences (Ext_SEQ): Multiple Ele_SEQ (both with and without FFT) are concatenated to form an Ext_SEQ. This allows for a richer representation of the protein, combining various physicochemical properties and spectral information.
  4. Modeling: A Partial Least Squares Regression (PLSR) algorithm is used to build predictive models based on these Ext_SEQ and experimentally determined protein activities.
  5. Validation: Models are evaluated using leave-one-out cross-validation (LOOCV), with performance metrics including the coefficient of determination (cvR²) and root mean squared error (cvRMSE).

The novelty lies in the systematic combination of multiple physicochemical descriptors with FFT-derived spectral information to create more informative input features for machine learning models.

Datasets

The authors evaluated their method on four publicly available datasets, each representing different protein lengths and activities:

  • GLP-2: 31 alanine variants, measuring potency (cAMP activation).
  • TNF alpha: 20 mutants and one wild-type, measuring relative affinity (%Ka) to its receptors.
  • Cytochrome P450: 184 chimeric sequences, measuring thermostability (T50).
  • Epoxide Hydrolase: 37 mutants and one wild-type, measuring enantioselectivity (E-value, transformed to ∆∆G‡).

Key Insights and Ideas

  • FFT Unlocks Hidden Information: Applying FFT to numerical protein sequences reveals spectral information that significantly improves prediction accuracy. This suggests that protein function is influenced by complex, non-linear interactions that are better captured in the frequency domain.
  • Combinatorial Power: Simply concatenating multiple physicochemical descriptors (Ext_SEQ) generally leads to better predictions than using a single descriptor. This highlights the benefit of a multi-faceted view of protein properties.
  • Descriptor Choice Matters: The optimal combination of descriptors and the use of FFT are protein- and fitness-dependent. There isn't a universal "best" set of descriptors; rather, the approach needs to be tailored to the specific problem.
  • Synthetic Descriptors Aren't Always Best: Surprisingly, established "synthetic" descriptors (like z-scales or principal components) that summarize multiple properties were not consistently the best performers. This suggests that carefully selected, individual physicochemical descriptors, especially when combined with spectral data, can be more effective.
  • Iterative Improvement: A "successive concatenation" approach, where the best descriptors are iteratively added, showed better performance than a purely combinatorial approach that explored all combinations.

Key Results

  • Significant Performance Gains: The innov'SAR method, particularly when using Ext_SEQ with FFT, demonstrated significant improvements in prediction accuracy across all four datasets. For example, for Cytochrome P450, cvR² improved from 0.83 (single index, no FFT) to 0.88 (three indices with FFT).
  • GLP-2 Example: Even with a limited training set (31 variants), the method achieved substantial improvements, showing its utility for datasets where experimental data is scarce.
  • Epoxide Hydrolase: While the base model was already excellent, combining indices still offered marginal improvements, suggesting that even highly accurate models can benefit from richer feature sets.
  • Computational Efficiency: The authors aimed for a method runnable on a personal computer, and the presented approaches, while computationally intensive, were shown to be manageable within reasonable timeframes.

Limitations and Future Work

  • Descriptor Selection: While the paper explores various combinations, the optimal selection of descriptors and FFT parameters remains somewhat empirical and dependent on the specific protein and fitness being studied.
  • Computational Cost: While manageable, generating and testing all possible Ext_SEQ combinations can still be computationally demanding, especially with a large number of indices. The authors explored limiting the number of indices and frequencies to mitigate this.
  • Generalizability of Classifications: The authors note that existing classifications of AAindex descriptors might not cover all relevant biochemical properties or biological activities. They suggest exploring new classification approaches, including statistical clustering.
  • Further Descriptor Exploration: The paper highlights the need for further investigation into descriptors, especially those derived from Fourier transformation, and their impact on modeling. This is slated for future work.
  • Broader Applicability: The authors believe their approach has broad interest for machine learning experts, biotechnologists, and molecular biologists, particularly for enhancing existing mutant libraries where screening has been challenging.

Summary updated

Abstract

The work aiming to unravel the correlation between protein sequence and function in the absence of structural information can be highly rewarding. We present a new way of considering descriptors from the amino acids index database for modeling and predicting the fitness value of a polypeptide chain. This approach includes the following steps: (i) Calculating Q elementary numerical sequences (Ele_SEQ) depending on the encoding of the amino acid residues, (ii) determining an extended numerical sequence (Ext_SEQ) by concatenating the Q elementary numerical sequences, wherein at least one elementary numerical sequence is a protein spectrum obtained by applying fast Fourier transformation (FFT), and (iii) predicting a value of fitness for polypeptide variants (train and/or validation set). These new descriptors were tested on four sets of proteins of different lengths (GLP-2, TNF alpha, cytochrome P450, and epoxide hydrolase) and activities (cAMP activation, binding affinity, thermostability and enantioselectivity). We show that the use of multiple physicochemical descriptors coupled with the implementation of the FFT, taking into account the interactions between residues of amino amides within the protein sequence, could lead to very significant improvement in the quality of models and predictions. The choice of the descriptor or of the combination of descriptors and/or FFT is dependent on the couple protein/fitness. This approach can provide potential users with value added to existing mutant libraries where screening efforts have so far been unsuccessful in finding improved polypeptide mutants for useful applications.
BibTeX
@article{fontaine2019novel,
  author = {Nicolas Fontaine and Xavier F. Cadet and Iyanar Vetrivel},
  title = {{Novel Descriptors and Digital Signal Processing- Based Method for Protein Sequence Activity Relationship Study}},
  journal = {International Journal of Molecular Sciences},
  year = {2019},
  doi = {10.3390/ijms20225640},
}
Figure