PaperlayerView exact revision on arXiv

Source-linked AI summary

HPSLPred: An Ensemble Multi-label Classifier for Human Protein Subcellular Location Prediction with Imbalanced Source

Shixiang Wan, Quan Zou

arXiv:1704.05204v1cs.LG

TL;DR

Protein localization experiments are costly, and existing prediction methods often use multi-class classification despite proteins occupying multiple locations and imbalanced labels. HPSLPred combines ensemble multi-label classification with approaches for imbalanced data. The paper reports higher accuracy and lower running time than state-of-the-art methods.

  • Problem

    Many existing methods treat localization as multi-class rather than multi-label prediction, and techniques for imbalanced multi-label data are scarce.

    Source: §1 · p. 1

  • Method

    HPSLPred is an ensemble multi-label classifier using compound features, imbalance handling, and optimal dimension searching.

    Source: §1 · p. 2, §5 · p. 11

  • Results

    The authors report higher accuracy and lower running time than state-of-the-art methods, and state that HPSLPred achieves the highest AP value.

    Source: §1 · p. 2, §5 · p. 11

  • Takeaways & Limitations

    A user-friendly webserver for HPSLPred is provided for users.

    Source: §5 · p. 11

    The authors identify a need for a more comprehensive, regularly updated protein database and more efficient algorithms.

    Source: §5 · p. 11

Abstract

from arXiv · show

Predicting the subcellular localization of proteins is an important and challenging problem. Traditional experimental approaches are often expensive and time-consuming. Consequently, a growing number of research efforts employ a series of machine learning approaches to predict the subcellular location of proteins. There are two main challenges among the state-of-the-art prediction methods. First, most of the existing techniques are designed to deal with multi-class rather than multi-label classification, which ignores connections between multiple labels. In reality, multiple locations of particular proteins implies that there are vital and unique biological significances that deserve special focus and cannot be ignored. Second, techniques for handling imbalanced data in multi-label classification problems are necessary, but never employed. For solving these two issues, we have developed an ensemble multi-label classifier called HPSLPred, which can be applied for multi-label classification with an imbalanced protein source. For convenience, a user-friendly webserver has been established at http://server.malab.cn/HPSLPred.

1. Introduction

Protein localization prediction matters for biological research, but experimental approaches are costly and existing computational methods often treat proteins as belonging to only one location. HPSLPred is proposed as an ensemble multi-label classifier for imbalanced human-protein data.

1. Introduction

  • Protein localization information is linked to genome annotation, function prediction, vaccine-target identification, and insights into protein interactions.The introduction also connects aberrant localization with disease.Source: §1 · p. 1
  • Experimental localization approaches are costly and time-consuming, motivating computational prediction as protein sequence resources grow.The cited passage reports 553,231 UniProtKB sequences as of 01-Dec-2016, compared with 3,939 in 1986.Source: §1 · p. 1
  • The paper describes prior methods as expanding the number of predicted sites and combining different sequence-feature approaches.Examples include amino acid composition, pseudo amino acid composition, and sequential evolution information.Source: §1 · p. 1
  • Most existing methods use multi-class prediction, although many proteins occupy multiple sites whose distinct biological roles merit attention.The introduction also identifies limited techniques for handling imbalanced data in multi-label classification.Source: §1 · p. 1
  • HPSLPred is proposed as an ensemble multi-label classifier for human protein localization with imbalanced source data.The approach is based on compound features and optimal dimension searching.Source: §1 · p. 2

2. Material and Methods

The methods construct a redundancy-filtered human protein dataset and address its multi-label imbalance by balancing each label-specific task. HPSLPred combines sequence features, dimension selection, and an ensemble of classifiers, with average precision used as the principal comparison metric.

2.1 Data construction with human protein

  • At a 0.7 CD-HIT similarity threshold, redundancy filtering yields 9,895 human protein sequences; about 30.60% of samples have multiple labels.The dataset retains sequences assigned to the selected locations, and the authors describe it as their final test dataset.Source: §2.1 · p. 3

2.2 Handling imbalanced source

  • The method converts the multi-label dataset into label-specific binary tasks, then uses SVM-based sampling near the decision boundary to form balanced datasets.The authors favor binary relevance because they regard the locations in their dataset as independent, while noting that balancing adds time complexity.Source: §2.2 · pp. 3–5

2.3 Features for subcellular localization

  • Protein sequences are represented using an 188-dimensional physicochemical feature model and seven Pse-in-One extraction methods.The 188 dimensions combine amino-acid proportions with eight groups of physicochemical properties; Pse-in-One includes autocorrelation and pseudo-amino-acid-composition methods.Source: §2.3 · p. 5

2.4 HPSLPred: a novel and ensemble multi-label classifier

  • HPSLPred combines the 188D and seven Pse-in-One feature methods into 350D features, applies MRMD reduction, and searches for an optimal dimension in two stages.The framework then evaluates an ensemble of scikit-learn classifiers on balanced single-label datasets; the reported per-label precision uses each label's maximum across the twelve classifiers.Source: §2.4 · pp. 5–6

2.5 Performance metrics

  • Average precision (AP), which ranges from 0 to 1, is the principal metric used for the comparison experiments.The paper also lists Hamming loss, Coverage, OneError, IsError, and Ranking loss as multi-label metrics.Source: §2.5 · pp. 6–7

3. Results

The experiments compare HPSLPred with Mulan and WEKA on 188D, PseAAC-26D, and hybrid 350D features, then compare it with state-of-the-art methods. HPSLPred reaches its reported highest AP, 75.89%, with the hybrid features and also leads the state-of-the-art comparison.

3. Results

  • The results section evaluates HPSLPred against Mulan and WEKA across 188D, PseAAC-26D, and hybrid 350D features, and compares it with state-of-the-art methods.The authors also report that source code is public and a user-friendly webserver is available.Source: §3 · p. 7

3.1 Comparative experiments based on 188D classical features

  • On 188D features, HPSLPred reaches 72.45% precision, exceeding Mulan and WEKA; each label’s precision is reported as higher than the other classifiers’ results.IBLR_ML has the highest precision among Mulan classifiers at 68.63%, while RandomForest leads WEKA at 68.20%.Source: Table 4 · p. 7, Figure 3 · p. 7, §3.1 · p. 8

3.2 Comparative experiments based on PseAAC-26D classical features

  • On PseAAC-26D features, HPSLPred achieves the highest AP, 74.21%, about 6% above the second-best algorithm.The authors report that Mulan with the multi-label source outperforms WEKA with balanced single-label datasets in this comparison.Source: Table 5 · p. 8, Figure 4 · p. 8, §3.2 · p. 8

3.3 Comparative experiments based on hybrid 350D features and dimensional searching

  • With hybrid 350D features and two-layer dimension searching, HPSLPred reaches the highest reported AP, 75.89%.The second search round examines dimensions from 20D to 30D; the authors report improvements over the 188D and PseAAC feature models.Source: Table 6 · p. 9, §3.3 · p. 9, Figure 6 · p. 10

3.4 Comparison with state-of-the-art methods

  • HPSLPred achieves the best AP, 75.89%, in the state-of-the-art comparison.The authors report that Xiaotong Guo’s approach performs comparably to Mulan, while Hum-mPLoc 2.0, mGOF-Loc, and IMMMLGP also have notable performance.Source: Table 7 · p. 10, §3.4 · p. 10

4. Discussion

The discussion reports HPSLPred’s AP results across feature models and classifier comparisons, and identifies redundant features in the hybrid model. It also finds no apparent relation between PseAAC-26D and 188D results.

  • Feature selection on the hybrid 350D model yielded the best AP, and HPSLPred performed convincingly in comparisons with state-of-the-art methods.The authors identify redundant expressions in the hybrid feature model as a reason feature selection was needed.Source: §4 · p. 10
  • On the 188D model, HPSLPred integrated twelve basic classifiers and grid parameter searching, reaching the highest AP among the compared classifiers.Source: §4 · p. 10
  • Mulan-based multi-label classifiers performed better than WEKA-based classical single-label classifiers in the reported comparison.Source: §4 · p. 10
  • The authors report that each label’s precision corresponded to its AP value, while PseAAC-26D results had no apparent relation to 188D results.They describe the latter as evidence of new expression in the two feature models.Source: §4 · p. 10

5. Conclusion

The conclusion describes HPSLPred’s multi-stage feature, balancing, and classifier pipeline for human protein subcellular-location prediction. It also identifies database expansion and more efficient classifiers as future work.

  • HPSLPred combines hybrid 350D features, BM-based conversion of the imbalanced multi-label source, SVM decision-boundary balancing, twelve classifiers, grid search, and MRMD-based dimension reduction to achieve the highest AP.The pipeline also uses multi-threading to accelerate dimensionality reduction.Source: §5 · p. 11
  • The authors established a user-friendly webserver for HPSLPred.Source: §5 · p. 11
  • Future work aims to build a more comprehensive, updated protein database and to use hardware and algorithm strategies for more efficient multi-label classifiers.The conclusion says multi-threading and GPU parallel technology reduce program time complexity and links more efficient classifiers with improving localization precision.Source: §5 · p. 11
Loading 1704.05204v1…