PaperlayerView exact revision on arXiv

Source-linked AI summary

Pretata: predicting TATA binding proteins with novel features and dimensionality reduction strategy

Quan Zou, Shixiang Wan, Ying Ju, Jijun Tang, Xiangxiang Zeng

arXiv:1703.02850v1q-bio.QMcs.LGq-bio.BM

TL;DR

Identifying TATA-binding proteins (TBPs) from sequence is important, and reducing feature dimensionality while maintaining prediction accuracy is a key challenge. PreTata combines novel sequence features with dimensionality search and reports over 90% TBP prediction accuracy, exceeding other methods.

Abstract

from arXiv · show

Background: It is necessary and essential to discovery protein function from the novel primary sequences. Wet lab experimental procedures are not only time-consuming, but also costly, so predicting protein structure and function reliably based only on amino acid sequence has significant value. TATA-binding protein (TBP) is a kind of DNA binding protein, which plays a key role in the transcription regulation. Our study proposed an automatic approach for identifying TATA-binding proteins efficiently, accurately, and conveniently. This method would guide for the special protein identification with computational intelligence strategies. Results: Firstly, we proposed novel fingerprint features for TBP based on pseudo amino acid composition, physicochemical properties, and secondary structure. Secondly, hierarchical features dimensionality reduction strategies were employed to improve the performance furthermore. Currently, Pretata achieves 92.92% TATA- binding protein prediction accuracy, which is better than all other existing methods. Conclusions: The experiments demonstrate that our method could greatly improve the prediction accuracy and speed, thus allowing large-scale NGS data prediction to be practical. A web server is developed to facilitate the other researchers, which can be accessed at http://server.malab.cn/preTata/.

Background

TBP binds DNA at TATA boxes and contributes to transcription, but computational TBP classification had not been studied. The paper proposes novel sequence features and undersampling with dimensionality search to address this task.

Background

  • TBP binds TATA boxes, helps DNA melting, and is an RNA polymerase component, making its role in gene regulation and health and disease important.Despite this biological importance, no studies had focused on computational TBP classification or prediction.Source: Background · p. 1
  • Protein-classification fingerprints include PseACC, K-mer and K-ship frequencies, composition and physicochemical features, secondary-structure features, and PSSM; feature selection and reduction are also used.Examples of reduction techniques include mRMR, t-SNE, and MRMD.Source: Background · pp. 1–2
  • The available TBP and non-TBP sequences formed an extremely imbalanced training set, risking overfitting and unsatisfactory confusion matrices despite high accuracy.The large number of negative sequences relative to positives also dramatically reduces algorithm efficiency.Source: Background · p. 2

Methods · Features based on composition and physicochemical properties of amino acids

The methods combine amino-acid composition and physicochemical descriptors with predicted secondary-structure features, then use MRMD and a two-stage search to reduce dimensionality. They also construct a negative sample set and use an SVM classifier for TBP prediction.

Methods · Features based on composition and physicochemical properties of amino acids

  • The 188D composition/physicochemical representation combines 20 amino-acid proportions with CTD descriptors for eight physicochemical properties.CTD captures composition, transitions between amino acids with different hydrophobic properties, and positions at the first, 25%, 50%, 75%, and last occurrences; each sequence yields 188 values.Source: Features based on composition and… · p. 2

Methods · Features from secondary structure

  • Secondary-structure features use PSIPRED predictions encoded as H, E, and C, from which global and local features form a 473D representation.PSIPRED uses an artificial neural network and PSI-BLAST alignment results.Source: Features from secondary structure · p. 3

Methods · Features dimensionality reduction

  • MRMD reduces feature dimensionality using feature–class correlation and inter-feature redundancy, then searches for an optimal dimension based on accuracy.It uses Pearson’s correlation coefficient for feature–class association and a distance function for redundancy.Source: Features dimensionality reduction · p. 3
  • The optimal-dimension search uses a coarse primary scan followed by a finer secondary scan to select the final dimension.The primary step searches a broad range quickly; the secondary step scans dimensions sequentially to determine final accuracy, sensitivity, and specificity.Source: Features dimensionality reduction · p. 4

Methods · Negative samples collection

  • Negative controls were sampled from proteins outside the positive sequences’ PFAMs, then iteratively updated with misclassified negatives near the classification boundary.At a 90% CD-HIT clustering threshold, the dataset contained 559 positive and 8,465 negative instances; 559 negatives were randomly sampled as controls.Source: Negative samples collection · p. 5

Methods · Support vector machine (SVM)

  • The study selected SVM over random forest, KNN, C4.5, and libD3C; ensemble learning performed as well as SVM but took much longer to run.The implementation used LIBSVM with an RBF kernel and grid-optimized parameters g = 0.5 and c = 128.Source: Support vector machine (SVM) · p. 5

Results

PreTata’s LIBSVM dimensionality search achieved the strongest reported results, while combined composition/physicochemical and secondary-structure features outperformed either feature set alone. SVM-selected negative samples did not improve performance, and PreTata outperformed BLASTP and PSI-BLAST in the reported comparison.

Measurements

  • The experiments assessed individual-class sensitivity (SN) and specificity (SP), alongside overall accuracy (ACC).Source: Measurements · p. 5

Joint features outperform the single ones

Dimensionality reduction outperforms the joint features

Negative samples have been highly representive

  • 84.05% ACC, 88.90% SN, and 79.20% SP with SVM-selected negatives did not improve on the random-negative experiment.The new negative set comprised support-vector negatives; the dataset contained 559 positive and 7908 negative sequences, with the longest-distance 20% of negatives reserved.Source: Negative samples have been highly… · pp. 8–9

Comparing with state-of-arts software tools

Discussion

PreTata addresses the need for accurate, efficient TBP prediction as NGS datasets grow, and its sequence fingerprints may guide detection of other special proteins from primary sequences.

  • PreTata quickly classifies and predicts TBPs from imbalanced datasets, addressing demand for an accurate, efficient platform as NGS research datasets grow.The authors describe continued improvement as a way to facilitate theoretical prediction.Source: Discussion · pp. 9–10
  • The method combines advanced machine learning with novel protein-sequence fingerprints to facilitate TBP identification.Source: Discussion · p. 10
  • The proposed fingerprints may also guide detection of other special proteins from primary sequences.Source: Discussion · p. 10

Conclusions

PreTata combines feature extraction with a secondary dimensionality search to identify TBPs, achieving over 90% accuracy and supporting practical large-scale NGS prediction.

Conclusions

  • The study describes feature extraction based on physicochemical properties and PSIPRED secondary-structure predictions, and PreTata uses a secondary dimensionality search.The physicochemical-property features are 188D, and the PSIPRED-based features are 473D.Source: Conclusions · p. 10
  • Over 90% accuracy in TBP prediction, with better accuracy than other methods and improved efficiency for practical large-scale NGS prediction.The authors report that experiments demonstrated the method’s effectiveness.Source: Conclusions · p. 10
  • An online web server and open-source software supporting massive data processing were developed to facilitate use of the method.The project is freely accessible at http://server.malab.cn/preTata/.Source: Conclusions · p. 10

Declarations

The article was published in BMC Systems Biology Volume 10, Supplement 4 (2016), the proceedings of the 27th International Conference on Genome Informatics: systems biology.

Declarations

  • The article appeared in BMC Systems Biology Volume 10, Supplement 4 (2016), whose full contents are available online.The supplement contains the proceedings of the 27th International Conference on Genome Informatics: systems biology.Source: Declarations · p. 10

Funding · Availability of data and materials

The work and publication costs were funded by the Natural Science Foundation of China; all generated and analyzed data are available in the published article and on the websites.

Funding

  • The Natural Science Foundation of China funded the work and publication costs (No. 61370010).Source: Funding · p. 10

Availability of data and materials

Authors’ contributions · Competing interests

The authors describe each contributor’s role in the project and manuscript, and declare no competing interests.

Authors’ contributions

Competing interests

Consent for publication · Ethics approval and consent to participate · Author details

The supplied material lists four author-affiliation institutions and gives the paper’s publication date.

Author details

  • Authors are affiliated with Tianjin University, Harbin Institute of Technology Shenzhen Graduate School, Xiamen University, and the University of South Carolina; the paper was published 23 December 2016.Source: Author details · p. 11
Loading 1703.02850v1…