Source-linked AI summary
SBSM-Pro: Support Bio-sequence Machine for Proteins
Yizheng Wang, Yixiao Zhai, Yijie Ding, Quan Zou
TL;DR
Protein classification requires computational methods that can interpret abundant biological sequence data and support experimental investigation. SBSM-Pro groups amino acids by physicochemical properties, aligns sequences to measure similarity, and combines kernels with MKL for SVM-based prediction. The model reports strong performance across ten datasets addressing protein function and posttranslational modification, while its authors identify usability and additional structural or functional kernels as future directions.
Problem
Protein classification needs computational support because large sequence collections require annotation, while biological sequence mechanisms can be difficult to interpret through individual experiments.
Method
SBSM-Pro groups amino acids by physicochemical properties, uses sequence alignment to construct similarity kernels, combines them with MKL, and applies an SVM for prediction.
Results
SBSM-Pro reports outstanding performance across ten datasets for identifying protein function and posttranslational modification.
Takeaways & Limitations
The framework supports protein-sequence classification by retaining information from raw sequences and integrating similarity kernels from different perspectives.
Takeaways & Limitations
The authors identify improving usability and adding structural and functional kernels as open directions for SBSM-Pro.
Abstract
from arXiv · showhide
Proteins play a pivotal role in biological systems. The use of machine learning algorithms for protein classification can assist and even guide biological experiments, offering crucial insights for biotechnological applications. We introduce the Support Bio-Sequence Machine for Proteins (SBSM-Pro), a model purpose-built for the classification of biological sequences. This model starts with raw sequences and groups amino acids based on their physicochemical properties. It incorporates sequence alignment to measure the similarities between proteins and uses a novel multiple kernel learning (MKL) approach to integrate various types of information, utilizing support vector machines for classification prediction. The results indicate that our model demonstrates commendable performance across ten datasets in terms of the identification of protein function and posttranslational modification. This research not only exemplifies state-of-the-art work in protein classification but also paves avenues for new directions in this domain, representing a beneficial endeavor in the development of platforms tailored for the classification of biological sequences. SBSM-Pro is available for access at http://lab.malab.cn/soft/SBSM-Pro/.
1 Introduction
Bio-sequence classification is increasingly important for annotating large sequence collections and guiding biological experiments, but existing approaches involve trade-offs between feature engineering and data-intensive deep learning. SBSM-Pro addresses these challenges by combining sequence-based kernels, physicochemical sequence processing, and SVM classification.
- 1 Introduction: Large-scale genome mapping has created an urgent need for computer-assisted annotation of sequence functions and multiple computational predictions to guide wet-lab exploration.Sequence variations are linked to complex disease mechanisms that cannot be fully addressed through one biological experiment.
- 1 Introduction: Bio-sequence classification commonly uses either extracted numerical features with traditional classifiers or direct sequence classification with deep learning.Relevant features include frequency, physicochemical, structural, and evolutionary information.
- 1 Introduction: Deep learning methods can classify directly from encoded sequences, but they require substantial training data and complex software that may reduce platform usability.These requirements may discourage researchers without information-science backgrounds.
- 1 Introduction: SBSM-Pro uses SVMs, sequence alignment, and an end-to-end kernel method to classify biological sequences without deep learning’s extensive data requirements.The method replaces numerical vectors with biological sequences and targets sequence structure and function prediction.
- 1 Introduction: The proposed PSD process reduces the amino acid alphabet using physicochemical properties, while MKL combines global and local sequence-similarity kernels for SVM prediction.The contributions include Levenshtein distance, Smith–Waterman scores, and a precomputed-kernel SVM.
2 Results and Discussion
Across ten protein classification datasets, SBSM-Pro generally outperformed existing methods while combining sequence grouping, alignment-based similarity, and multiple-kernel learning. Results also show that grouping benefits functional datasets more clearly, SW alignment generally exceeds LS distance, and dictionary suitability varies by dataset.
- Comparative performance: 0.1853 higher ACC was achieved by SBSM-Pro than the prior SVM-based DBP method.The earlier method emphasized evolutionary features exclusively, whereas SBSM-Pro used the original protein sequence.
- Comparative performance: Approximately 3.98% and 5.10% ACC improvements were obtained on PVP and PTSS, respectively, over the best existing methods.The compared methods primarily used amino acid composition information and were reported to suffer from overfitting.
- Comparative performance: 16.29% higher ACC was achieved by SBSM-Pro than iGlu AdaBoost on the PLGS dataset.iGlu AdaBoost used feature representations, selection, and AdaBoost classification, with potential generalizability concerns noted.
- Comparative performance: SBSM-Pro generally outperformed existing methods across 10 commonly used amino acid classification datasets.The evaluation used ACC and compared SBSM-Pro with leading contemporary models across ten datasets.
- Dictionary and alignment effects: Amino acid grouping substantially improved ACC on DBP, T3SE, and PVP, but its enhancement was less pronounced across seven PTM datasets.The authors associate the weaker PTM effect with relatively shorter protein sequences, which reduce the noise-reduction benefit of alphabet reduction.
- Dictionary and alignment effects: SW alignment consistently outperformed LS distance, while dictionary performance varied across datasets because physicochemical properties contribute differently to classification tasks.SW can insert gaps during alignment, producing better alignments and more accurate sequence-similarity representations; MKL integrates kernels from all dictionaries with different weights.
3 Conclusion
SBSM-Pro achieved strong results across multiple protein-classification datasets, with ablation studies supporting the effectiveness of its modules. The approach combines sequence-based kernels, amino acid grouping, and MKL, while future work targets broader protein representations and DNA/RNA classification.
- SBSM-Pro achieved outstanding results across multiple datasets, and ablation studies demonstrated the effectiveness and indispensability of its modules.
- PSD reduces the amino acid alphabet to improve sequence alignment and represent protein-sequence similarities more accurately.
- SBSM-Pro uses Levenshtein distance and Smith–Waterman scoring to compute protein similarities before SVM classification.
- The method extracts multifaceted information from raw protein sequences and achieves higher accuracy than traditional feature-extraction techniques.
- HCKDM-MKL integrates sequence similarity kernels from different amino acid dictionaries by assigning them different weights.
- Future work includes structural and functional kernels, a graphical user interface, and extending classification beyond proteins to DNA and RNA sequences.
4 Materials and Methods
SBSM-Pro processes protein datasets and amino acid physicochemical data, constructs central kernels, and fuses them with an MKL strategy for final classification.
- The workflow begins by collecting and processing protein datasets into labeled samples, then retrieving and preprocessing amino acid physicochemical data.
- An MKL strategy fuses central kernels, and the resulting fused kernel is supplied to SBSM-Pro for classification.
4.1 Datasets
The study evaluates SBSM-Pro on ten protein-classification datasets covering protein functions and posttranslational modifications, using controlled partitions or cross-validation for fair comparison.
- The evaluation uses 10 datasets: 3 for protein-function identification and 7 for posttranslational-modification identification.
- The datasets cover functionalities including DNA-binding proteins, type III secreted effectors, and phage virion proteins.
- Protein-function datasets have varying sequence lengths, whereas sequence lengths are consistent within each PTM-identification dataset.
- The study uses identical training and testing sets when comparing algorithms to remove variation caused by different data partitions.
- Cross-validation partitions data into K subsets, evaluates each subset as testing data once, and averages the K resulting outcomes.
4.2 Physicochemical properties of amino acids
The method compiles amino acid physicochemical properties and processes them into grouping dictionaries through spectral clustering, while addressing incomplete property data.
- Amino acids are the basic components of proteins and possess diverse physicochemical properties arising from their molecular structures.
- The workflow collects ten commonly used datasets and produces corresponding labels and samples for analysis.
- Protein sequence-length distributions differ across datasets, with PTM datasets showing identical lengths within each dataset.
- The collected amino acid properties include alphacarbon positions, hydrophobicity, secondary structure, non-bonded energy, membrane regions, polarity and bulkiness, chemical structure, mean polarities, and side-chain properties.
- The properties are numerically represented and used to generate grouping dictionaries through spectral clustering, but incomplete data require further processing.
- Missing not-applicable values can arise because simple amino acid side chains contribute less pronounced signals in detailed side-chain analysis.
4.3 Amino acid grouping
Amino acids are grouped by physicochemical similarity through spectral clustering and represented as compact dictionaries. The resulting PSD process reduces the sequence alphabet, limits substitution noise, and supports protein sequence alignment.
- Amino acid grouping: PSD groups amino acids by physicochemical properties using spectral clustering, then constructs dictionaries that encode each group.Each original amino acid residue is replaced by its group number to produce a re-encoded protein sequence.
- Spectral clustering: Spectral clustering constructs a similarity matrix from Gaussian-kernel Euclidean distances before Laplacian eigendecomposition and eigenvector-based partitioning.The number of clusters is selected using the Calinski–Harabasz index, which evaluates within-cluster compactness and between-cluster separation.
- Spectral clustering: The method clusters the 20 amino acids despite the unknown appropriate cluster count.Candidate cluster numbers are evaluated because too few clusters lose sequence information, whereas too many fail to reduce the amino acid alphabet effectively.
- Dictionaries for grouping: PSD reduces the original protein alphabet and can reduce alignment noise from substitutions with little or no functional effect.The process also links sequence information with physicochemical properties to represent interprotein distances more accurately.
4.4 Generating Sequence Kernels
SBSM-Pro derives protein similarity kernels from sequence alignment. It uses Smith–Waterman scores for local similarity and Levenshtein distances for sequence dissimilarity, followed by normalization and centralization.
- Kernel construction: Re-encoded protein sequences are aligned to generate Smith–Waterman scores and Levenshtein distances, which are transformed into sequence similarity kernels.Normalization and centralization are applied after alignment-based measurements.
- Smith–Waterman kernel: Smith–Waterman identifies optimal local alignments by dynamic programming and traces back from the highest scoring matrix value until reaching zero.In the illustrated example, a seven-amino-acid local alignment produces a final SW score of 4.
- Smith–Waterman kernel: Pairwise Smith–Waterman scores are stored in a symmetric matrix KSW to form a protein similarity kernel.The matrix contains SW scores computed between all sequence pairs in the sample set.
- Levenshtein kernel: Levenshtein distance measures dissimilarity as the minimum number of insertions, deletions, and substitutions needed to transform one sequence into another.The distances are normalized to obtain the LS protein similarity kernel KLS.
4.5 Multiple kernel learning
The method combines multiple sequence-derived kernels through HCKDM-MKL. It measures global and local kernel dependence, balances them with λ, and optimizes a fused kernel for classification.
- Kernel fusion: Multiple kernel learning combines LS and SW kernels to capture complementary information from different protein-similarity representations.The kernel matrices are weighted through coefficients β to form a fused representation.
- Dependence measurement: HSIC evaluates dependence between kernel and label matrices, with higher values indicating stronger dependence and zero indicating independence or irrelevance.The method uses centered kernel operations and trace-based calculations to estimate HSIC.
- Kernel preprocessing: Kernel centering subtracts row and column means so similarity information is emphasized before MKL processing.The centered kernel is denoted bKp, while Kp denotes the original pth similarity kernel.
- Local dependence: Local kernels compactly represent specific sample characteristics and reduce computational resource use during kernel dependence measurement.HCKDM extracts local kernels from groups of K-nearest neighbors before applying HSIC.
- Hybrid dependence: HCKDM-MKL jointly measures global and local kernel dependence using λ, where 0 ≤ λ ≤ 1 controls the trade-off between the two alignments.The fused kernel is obtained by maximizing this hybrid dependence in an optimization problem.
4.6 Support vector machine with precomputed kernel
SBSM-Pro parameterizes the fused similarity kernel for compatibility with an SVM. The resulting dual optimization uses support-sequence similarities and standard label and box constraints.
- Kernel parameterization: The MKL-derived similarity kernel is parameterized so it can be used by the SVM without increasing computational complexity.Parameterization maps data into a higher-dimensional feature space and supports nonlinear relationships.
- SVM optimization: The SVM dual objective maximizes a term involving Lagrange multipliers, labels, and pairwise fused-kernel similarities.The fused kernel matrix K̂* corresponds to the kernel function K̂*(xi, xj).
- SVM constraints: The equality constraint requires the sum of αi yi over training samples to equal zero.The box constraint keeps each αi nonnegative and no larger than C, which controls the margin-error trade-off.
Appendix A Supplementary material
Appendix A provides supplementary tables describing processed physicochemical properties, amino-acid grouping dictionaries, and dataset sources.
- Table A1 summarizes the processed physicochemical properties of amino acids.
- The dataset is obtained from two cited studies.
- Tables A2–A4 provide dictionaries for amino-acid groupings D1, D2, and D3.