PaperlayerView exact revision on arXiv

Source-linked AI summary

An adaptively weighted statistic for detecting differential gene expression when combining multiple transcriptomic studies

Jia Li, George C. Tseng

arXiv:1108.3180v2stat.AP

TL;DR

Combining gene-expression studies is challenging because common list-based approaches can be too conservative or insufficiently conservative, and existing methods lacked comprehensive evaluation under rigorously formulated hypotheses. The paper proposes an adaptively weighted statistic and compares it with established methods; AW performs near the best across contrasting alternatives and its weights categorize detected genes for further investigation.

  • Problem

    Combining results across gene-expression studies is difficult, and list intersections or unions can be too conservative or insufficiently conservative; the authors also note a lack of comprehensive method evaluations tied to rigorously formulated hypotheses.

    Source: §1 · pp. 2, 4

  • Method

    The authors propose an adaptively weighted statistic for combining multiple studies’ evidence, with weights that reflect each study’s contribution to a gene’s statistical significance.

    Source: §3 · pp. 8, 10

  • Results

    AW performed close to the best method at either extreme alternative setting, while EW was more powerful when all data sets were significant and minP when only one or a few were significant.

    Source: §6 · p. 21

  • Takeaways & Limitations

    AW weights can categorize differentially expressed genes and support further biological exploration, while concordance filtering can exclude discordantly regulated genes.

    Source: §3 · p. 8, §6 · p. 21

    The authors caution that AW is inadequate for traditional epidemiological or evidence-based medical meta-analysis because its selection procedure favors studies with concordant significant effects.

    Source: §3 · p. 8

Abstract

from arXiv · show

Global expression analyses using microarray technologies are becoming more common in genomic research, therefore, new statistical challenges associated with combining information from multiple studies must be addressed. In this paper we will describe our proposal for an adaptively weighted (AW) statistic to combine multiple genomic studies for detecting differentially expressed genes. We will also present our results from comparisons of our proposed AW statistic to Fisher's equally weighted (EW), Tippett's minimum p-value (minP) and Pearson's (PR) statistics. Due to the absence of a uniformly powerful test, we used a simplified Gaussian scenario to compare the four methods. Our AW statistic consistently produced the best or near-best power for a range of alternative hypotheses. AW-obtained weights also have the additional advantage of filtering discordant biomarkers and providing natural detected gene categories for further biological investigation. Here we will demonstrate the superior performance of our proposed AW statistic based on a mix of power analyses, simulations and applications using data sets for multi-tissue energy metabolism mouse, multi-lab prostate cancer and lung cancer.

1. Introduction.

Combining genomic studies is increasingly important, but differences in data structures and hypotheses can make effect-size integration difficult. The paper motivates testing methods for combining study-level significance and proposes an adaptively weighted approach for detecting differentially expressed genes.

1. Introduction.

  • Differences in data structures and statistical hypotheses can make direct combination of effect sizes difficult or impossible across studies.Combining transformed p-values is more feasible because it depends on individual tests rather than underlying data structures.Source: §1 · pp. 1–2
  • Significance-combination methods include Fisher’s equally weighted statistic, Tippett’s minimum p-value, and other transformations, while effect-size methods include random-effects and Bayesian approaches.When studies have similar designs and measure outcomes similarly, combining effect sizes is usually preferred to combining significance.Source: §1 · pp. 2–3
  • Microarray meta-analysis based on intersecting or unioning gene lists can be too conservative or insufficiently conservative, respectively, especially as the number of studies grows.The paper distinguishes this practice from statistical methods that combine study-level evidence.Source: §1 · p. 2
  • The paper proposes an adaptively weighted statistic and compares it with commonly used methods using simulations, genomic-study applications, and Gaussian power analyses.The applications include mouse energy-metabolism, prostate-cancer, and lung-cancer data sets.Source: §1 · pp. 3–4
  • Existing meta-analysis methods had not been comprehensively evaluated, which the authors attribute primarily to a lack of rigorous statistical hypothesis formulation.The paper considers gene expression profiles from K studies.Source: §1 · p. 4

2. Two major complementary hypothesis settings.

The section distinguishes study-invariant biomarkers from biomarkers that are differentially expressed in only some studies. The authors use these complementary hypotheses to frame biomarker targets and motivate focusing on the study-specific alternative.

2. Two major complementary hypothesis settings.

  • Under H_A, a gene is selected only if it is differentially expressed in all studies; under H_B, expression in one or more studies is sufficient.The authors treat H_A and H_B as complementary interpretations for applications, with H_A contained in H_B.Source: §2 · p. 5
  • For prostate-cancer studies, study-invariant biomarkers are of greater biological interest because study-specific findings may reflect technical discrepancies or sample-population heterogeneity.The authors note that further investigation of study-specific biomarkers could provide technical insights without necessarily providing biological insights into the disease.Source: §2 · p. 5
  • The mouse example distinguishes tissue-invariant targets from tissue-specific signatures, while the section also describes prostate- and lung-cancer study sets.The mouse studies compare brown fat, liver, and heart across wild-type and VLCAD-deficient genotypes.Source: §2 · p. 5, Table 1 · p. 6
  • Methods target different hypothesis spaces and biomarker lists; for example, Bayesian PI and SEI tend to detect H_A-type biomarkers, while EW and minP have different power patterns across H_B alternatives.PI performs better than SEI for genes with a high mean effect in one study and a low mean effect in another; EW tends to be more powerful when h is large, and minP when h is small.Source: §2 · p. 7, Table 2 · p. 7
  • The authors focus on H_B and present AW as robust near either extreme of the number of affected studies, with examples of better performance than EW and minP in intermediate scenarios.AW can distinguish study-invariant genes from study-specific genes and indicate which studies show differential expression.Source: §2 · pp. 7–8

3. Adaptively-weighted statistic.

Because biomarkers can change in study-specific ways, the section defines AW as choosing the weighted statistic with the smallest p-value over a prespecified weight space.

3. Adaptively-weighted statistic.

  • AW defines each gene's statistic as the smallest p-value across candidate weights, using a binary weight space with O(2^K −1) computation for K ≤10.The selected weight vector also provides a basis for gene categorization in follow-up biological interpretation.Source: §3 · p. 8
  • The binary search was chosen for computational convenience and biological interpretability; AW is inadequate for traditional epidemiological meta-analysis because selection favors concordant significant effects.For genomic screening, the selected weights support gene categorization, while AW data can distinguish similar or dissimilar expression across tissue pairs.Source: §3 · p. 8
  • Although uniform null p-values would yield a Gamma reference distribution, the procedure uses permutation tests and controls FDR at 5%.The section notes that uniform p-values are usually not true in real applications.Source: §3 · pp. 8–10
  • Study-wise inputs are penalized t-statistics, with s0 set to median genome variability, and permutation-derived p-values.The method permutes group labels within each study to calculate the permuted statistics and estimate gene-level p-values.Source: §3 · pp. 9, 11
  • An added step splits detected genes into concordant and discordant lists, which can be examined for biological discrepancies or annotation errors.EW and minP lack a direct criterion for this split; PR does not identify which studies distinguish concordant from discordant patterns.Source: §3 · p. 10
  • The general framework allows alternative test statistics and rejection regions, including an F-statistic for studies with multiple sample groups.The additional concordance-filtering step is presented for two-group comparisons using moderated t-statistics.Source: §3 · p. 10

4. Applications.

Across simulations and three applications, the adaptively weighted (AW) method stayed near the scenario-specific power leader and grouped detected genes by study or tissue patterns. In the applications, its gene lists were generally more interpretable and easier to assess for concordance than lists from competing methods.

4.1. Simulation study.

  • At θ = 2, minP detected 41.6 genes, EW 7.6, and AW 32.1; at θ = 1.5, EW detected 386.8, minP 121.3, and AW 359.3.AW performance remained stable in these extreme scenarios; most methods had FDR close to 5%, while maxP lost substantial power in scenario 1.Source: Table 3 · p. 11, §4.1 · pp. 11–12

4.2. Energy metabolism in mouse model.

  • In the mouse data, AW detected 203 genes, including 28 with conflicting regulation; its 175-gene concordance-filtered set exposed tissue patterns, while EW detected 329 harder-to-interpret genes.AW weights grouped 55 genes across all three tissues and 27 in liver and heart but not brown fat; the authors report more heart-related genes than brown-fat- or liver-related genes.Source: Figure 1 · p. 14, §4.2 · p. 15
  • AW’s weights can identify a gene with moderate brown-fat and liver signals and a very strong heart signal as significant across all three studies.The authors argue that combining the signals supports classifying this gene as a (1,1,1)-type gene rather than ignoring the weaker tissue signals after multiple-comparison adjustment.Source: §4.2 · p. 15

4.3. Prostate cancer and lung cancer studies.

  • For prostate cancer, AW found 722 genes, of which 618 were concordant; 14% were discordant, and 130 concordant genes were (1,1,1)-type biomarkers.EW, minP, and PR detected 924, 745, and 882 biomarkers, respectively, but their lists were harder to interpret and did not guarantee concordance.Source: §4.3 · p. 16, Figure 2 · p. 17
  • For lung cancer, AW detected 366 genes, of which 349 were concordant and 4.6% discordant, compared with 14.4% discordance in prostate cancer.Among concordant lung biomarkers, 99 (28.4%) were (1,1,1)-type and 96 (27.5%) were single-study-specific; the authors report more consistent regulation patterns and fewer single-study-specific biomarkers than in prostate cancer.Source: §4.3 · p. 16, Figure 3 · p. 18

5. Power and admissibility.

The section compares combined-test power in a simplified Gaussian setting and assesses admissibility. AW performs stably near the strongest method across sparse and widespread alternatives.

5. Power and admissibility.

  • In the simplified Gaussian setting, AW, EW, PR, and minP are admissible, whereas maxP is not.The analysis compares two-sample means with known variances and uses two-sided p-values.Source: §5 · p. 19

5.1. Power comparison of EW, AW, minP, maxP and PR under HBh′.

  • When h is small, minP is more powerful than EW; when h is large, EW is more powerful, while AW performs stably near the best method at both extremes.The comparison uses K = 10 studies, with power curves simulated 10,000 times for θ ∈ {1.2, 1.4}.Source: §5.1 · p. 20
  • The power comparison varies the number h of studies with nonzero effects, assigning them a common effect size and evaluating power as h changes.The stated setup uses σk = 1 and effect sizes θk to represent departures from the null.Source: §5.1 · p. 20

6. Discussion.

The discussion presents AW as a compromise between EW and minP, with weights that categorize detected genes for biological follow-up. It also notes constraints on the weight search, data setting, and scope of application.

6. Discussion.

  • EW is more powerful when all data sets are significant, minP when only one or a few are significant, and AW is close to the best at either extreme.Simulations support this robust pattern; AW weights also categorize genes for biological exploration, and a modified algorithm targets nondiscordant genes.Source: §6 · p. 21
  • The binary 0,1 weight search was chosen for computational convenience and biological interpretability; nonbinary weights may be slightly superior statistically, but not biologically.Nonbinary weights can produce biomarker lists similar to EW.Source: §6 · pp. 21–22
  • The analysis assumes identical matched gene lists with no missing values and focuses on two-group comparisons; missing-value extensions are needed to retain genes absent from some studies.For more than two groups, F-statistics can be combined, but small p-values alone do not ensure concordant expression patterns.Source: §6 · p. 22
  • The proposed method focuses on HSB rather than HSA, which the discussion notes is not the case in many biological applications.The paper also identifies gene-set enrichment as a possible extension of AW.Source: §6 · p. 23
  • Although the study combines microarray data, the methods are described as extendable to genomic, epigenomic, and proteomic studies; additional models may be needed for biological knowledge and specific questions.Examples include SNP, genome, and methylation arrays, proteomic experiments, and ChIP-on-chip experiments.Source: §6 · p. 23

APPENDIX: ADMISSIBILITY

The appendix connects admissibility to convex acceptance regions under the stated conditions. It uses this criterion to establish AW's admissibility and contrast combined tests.

APPENDIX: ADMISSIBILITY

  • Under HB with an exponential-family test statistic, a combined test is admissible exactly when its acceptance region is convex.The appendix applies this criterion to acceptance regions for the combined statistics.Source: APPENDIX: ADMISSIBILITY · p. 23
  • EW and minP are admissible, maxP is not, and PR is admissible; PR's status follows Owen's clarification of Birnbaum's interpretation.The appendix also notes that PR favors common directions of effects across studies.Source: APPENDIX: ADMISSIBILITY · p. 23
  • AW is admissible under HB and assumption (5.1) because its acceptance region is convex.The proof builds the region from convex sets and uses that their intersection is convex.Source: APPENDIX: ADMISSIBILITY · p. 24
Loading 1108.3180v2…