Wilson Wen Bin Goh

Fuzzy-FishNET: a highly reproducible protein complex-based approach for feature selection in comparative proteomics

BackgroundThe hypergeometric enrichment analysis approach typically fares poorly in feature-selection stability due to its upstream reliance on the t-test to generate differential protein lists before testing for enrichment on a protein complex, subnetwork or gene group.MethodsSwapping the t-test in favour of a fuzzy rank-based weight system similar to that used in network-based methods like Quantitative Proteomics Signature Profiling (QPSP), Fuzzy SubNets (FSNET) and paired FSNET (PFSNET) produces dramatic improvements.ResultsThis approach, Fuzzy-FishNET, exhibits high precision-recall over three sets of simulated data (with simulated protein complexes) while excelling in feature-selection reproducibility on real data (based on evaluation with real protein complexes). Overlap comparisons with PFSNET shows Fuzzy-FishNET selects the most significant complexes, which are also strongly class-discriminative. Cross-validation further demonstrates Fuzzy-FishNET selects class-relevant protein complexes.ConclusionsBased on evaluation with simulated and real datasets, Fuzzy-FishNET is a significant upgrade of the traditional hypergeometric enrichment approach and a powerful new entrant amongst comparative proteomics analysis methods.

Can Peripheral Blood-Derived Gene Expressions Characterize Individuals at Ultra-high Risk for Psychosis?

The ultra-high risk (UHR) state was originally conceived to identify individuals at imminent risk of developing psychosis. Although recent studies have suggested that most individuals designated UHR do not, they constitute a distinctive group, exhibiting cognitive and functional impairments alongside multiple psychiatric morbidities. UHR characterization using molecular markers may improve understanding, provide novel insight into pathophysiology, and perhaps improve psychosis prediction reliability. Whole-blood gene expressions from 56 UHR subjects and 28 healthy controls are checked for existence of a consistent gene expression profile (signature) underlying UHR, across a variety of normalization and heterogeneity-removal techniques, including simple log-conversion, quantile normalization, gene fuzzy scoring (GFS), and surrogate variable analysis. During functional analysis, consistent and reproducible identification of important genes depends largely on how data are normalized. Normalization techniques that address sample heterogeneity are superior. The best performer, the unsupervised GFS, produced a strong and concise 12-gene signature, enriched for psychosis-associated genes. Importantly, when applied on random subsets of data, classifiers built with GFS are “meaningful” in the sense that the classifier models built using genes selected after other forms of normalization do not outperform random ones, but GFS-derived classifiers do. Data normalization can present highly disparate interpretations on biological data. Comparative analysis has shown that GFS is efficient at preserving signals while eliminating noise. Using this, we demonstrate confidently that the UHR designation is well correlated with a distinct blood-based gene signature.

Overcoming analytical reliability issues in clinical proteomics using rank-based network approaches

Proteomics is poised to play critical roles in clinical research. However, due to limited coverage and high noise, integration with powerful analysis algorithms is necessary. In particular, network-based algorithms can improve selection of reproducible features in spite of incomplete proteome coverage, technical inconsistency or high inter-sample variability. We define analytical reliability on three benchmarks --- precision/recall rates, feature-selection stability and cross-validation accuracy. Using these, we demonstrate the insufficiencies of commonly used Student’s t-test and Hypergeometric enrichment. Given advances in sample sizes, quantitation accuracy and coverage, we are now able to introduce and evaluate Ranked-Based Network Approaches (RBNAs) for the first time in proteomics. These include SNET (SubNETwork), FSNET (FuzzySNET), PFSNET (PairedFSNET). We also introduce for the first time, PPFSNET(samplePairedPFSNET), which is a paired-sample variant of PFSNET. RBNAs (particularly PFSNET and PPFSNET) excelled on all three benchmarks and can make consistent and reproducible predictions even in the small-sample size scenario (n=4). Given these qualities, RBNAs represent an important advancement in network biology, and is expected to see practical usage, particularly in clinical biomarker and drug target prediction.

Valproic acid mediates miR-124 to down-regulate a novel protein target, GNAI1

Valproic acid (VPA) is an anti-convulsant drug that is recently shown to have neuroregenerative therapeutic actions. In this study, we investigate the underlying molecular mechanism of VPA and its effects on Bdnf transcription through microRNAs (miRNAs) and their corresponding target proteins. Using in silico algorithms, we predicted from our miRNA microarray and iTRAQ data that miR-124 is likely to target at guanine nucleotide binding protein alpha inhibitor 1 (GNAI1), an adenylate cyclase inhibitor. With the reduction of GNAI1 mediated by VPA, the cAMP is enhanced to increase Bdnf expression. The levels of GNAI1 protein and Bdnf mRNA can be manipulated with either miR-124 mimic or inhibitor. In summary, we have identified a novel molecular mechanism of VPA that induces miR-124 to repress GNAI1. The implication of miR-124→GNAI1→BDNF pathway with valproic acid treatment suggests that we could repurpose an old drug, valproic acid, as a clinical application to elevate neurotrophin levels in treating neurodegenerative diseases.

Dealing with Confounders in Omics Analysis

The Anna Karenina effect is a manifestation of the theory-practice gap that exists when theoretical statistics are applied on real-world data. In the course of analyzing biological data for differential features such as genes or proteins, it derives from the situation where the null hypothesis is rejected for extraneous reasons (or confounders), rather than because the alternative hypothesis is relevant to the disease phenotype. The mechanics of applying statistical tests therefore must address and resolve confounders. It is inadequate to simply rely on manipulating the P-value. We discuss three mechanistic elements (hypothesis statement construction, null distribution appropriateness, and test-statistic construction) and suggest how they can be designed to foil the Anna Karenina effect to select phenotypically relevant biological features.

Inverting proteomics analysis provides powerful insight into the peptide/protein conundrum

In proteomics, a large proportion of mass spectrometry (MS) data is ignored due to the lack of, or insufficient statistical evidence for mappable peptides. In reality, only a small fraction of features are expected to be differentially relevant anyway. Mapping spectra to peptides and subsequently, proteins, produces uncertainty at several levels. We propose it is better to analyze proteomic profiling data directly at MS level, and then relate these features to peptides/proteins. In a renal cancer data comprising 12 normal and 12 cancer subjects, we demonstrate that a simple rule-based binning approach can give rise to informative features. We note that the peptides associated with significant spectral bins gave rise to better class separation than the corresponding proteins, suggesting a loss of signal in the peptide-to-protein transition. Additionally, the binning approach sharpens focus on relevant protein splice forms rather than just canonical sequences. Taken together, the inverted raw spectra analysis paradigm, which is realised by the MZ-Bin method described in this article, provides new possibilities and insights, in how MS-data can be interpreted.

AI Paradigms for Teaching Biotechnology

Artificial intelligence (AI) is profoundly changing biotechnological innovation. Beyond direct application, it is also a useful tool for adaptive learning and forging new conceptual connections within the vast network of knowledge for the advancement of biotechnology. Here, we discuss a new paradigm for biotechnology education that involves coevolution with AI.

Turning straw into gold: building robustness into gene signature inference

Reproducible and generalizable gene signatures are essential for clinical deployment, but are hard to come by. The primary issue is insufficient mitigation of confounders: ensuring that hypotheses are appropriate, test statistics and null distributions are appropriate, and so on. To further improve robustness, additional good analytical practices (GAPs) are needed, namely: leveraging existing data and knowledge; careful and systematic evaluation of gene sets, even if they overlap with known sources of confounding; and rigorous testing of inferred signatures against as many published data sets as possible. Here, using a re-examination of a breast cancer data set and 48 published signatures, we illustrate the value of adopting these GAPs.

Why breast cancer signatures are no better than random signatures explained

Random signature superiority (RSS) occurs when random gene signatures outperform published and/or known signatures. Unlike reproducibility and generalizability issues, RSS is relatively underexplored. Yet, understanding it is imperative for better analytical outcome. In breast cancer, RSS correlates strongly with enrichment for proliferation genes and signature size. Removal of proliferation genes from random signatures reduces the predictive power of random signatures. Almost all genes are correlated to a certain extent with the proliferation signature, making complete elimination of its confounding effects impossible. RSS goes beyond breast cancer, because it also exists in other diseases; it is especially strong in other cancers in a platform-independent manner, and less severe, but present nonetheless, in nonproliferative diseases.

Fuzzy-FishNet: A highly precise distribution-free network approach for feature selection in clinical proteomics

Network-based analysis methods can help resolve coverage and inconsistency issues in proteomics data. Previously, it was demonstrated that a suite of rank-based network approaches (RBNAs) provides unparalleled consistency and reliable feature selection. However, reliance on the t-statistic/t-distribution and hypersensitivity (coupled to a relatively flat p-value distribution) makes feature prioritization for validation difficult. To address these concerns, a refinement based on the fuzzified Fisher exact test, Fuzzy-FishNet was developed. Fuzzy-FishNet is highly precise (providing probability values that allows exact ranking of features). Furthermore, feature ranks are stable, even in small sample size scenario. Comparison of features selected by genomics and proteomics data respectively revealed that in spite of relative feature stability, cross-platform overlaps are extremely limited, suggesting that networks may not be the answer towards bridging the proteomics-genomics divide.Network-based analysis methods can help resolve coverage and inconsistency issues in proteomics data. Previously, it was demonstrated that a suite of rank-based network approaches (RBNAs) provides unparalleled consistency and reliable feature selection. However, reliance on the t-statistic/t-distribution and hypersensitivity (coupled to a relatively flat p-value distribution) makes feature prioritization for validation difficult. To address these concerns, a refinement based on the fuzzified Fisher exact test, Fuzzy-FishNet was developed. Fuzzy-FishNet is highly precise (providing probability values that allows exact ranking of features). Furthermore, feature ranks are stable, even in small sample size scenario. Comparison of features selected by genomics and proteomics data respectively revealed that in spite of relative feature stability, cross-platform overlaps are extremely limited, suggesting that networks may not be the answer towards bridging the proteomics-genomics divide.