Interview Questions for Microarray Data Processing

Microarray technology is a powerful tool used to study gene expression on a large scale. Imagine you have thousands of different colored beads, each representing a specific gene. Microarrays essentially do the same thing, but instead of beads, they use microscopic spots of DNA (probes) attached to a solid surface, like a glass slide. Each spot represents a different gene. You then introduce labeled cDNA (complementary DNA) from your sample, which will bind to the complementary probes on the array. The amount of binding (measured by fluorescence intensity) indicates the level of expression of that specific gene in your sample. Higher fluorescence means higher gene expression.

Think of it like a detective searching a crime scene for clues. Each spot on the microarray is a potential clue, and the intensity of the fluorescence reveals the significance of that clue in solving the case (understanding gene expression).

There are several types of microarrays, each with its own advantages and disadvantages:

• cDNA microarrays: These are created by spotting cDNA clones onto a solid surface. They’re relatively inexpensive to produce but can suffer from cross-hybridization (where similar sequences bind to the wrong spots) and variations in spot quality.
• Oligonucleotide microarrays: These use short, synthetic DNA sequences (oligonucleotides) as probes. They offer higher specificity and more consistent spot quality than cDNA microarrays, making them better for detecting single nucleotide polymorphisms (SNPs), but are more expensive to produce.
• aRNA microarrays: These use amplified RNA as probes. This method provides higher sensitivity than cDNA microarrays.

The choice of microarray type depends on the specific research question, budget, and desired level of accuracy.

Microarray data preprocessing is crucial for removing artifacts and noise from the raw data, ensuring reliable analysis. The common steps typically include:

• Background correction: Subtracting non-specific binding signals from the measured intensities.
• Normalization: Adjusting for systematic variations between arrays or samples, such as differences in labeling efficiency or scanner settings.
• Filtering: Removing low-quality or unreliable data points, such as spots with low intensity or high variability.
• Transformation: Applying mathematical transformations (e.g., log transformation) to stabilize variance and improve data distribution for statistical analysis.

This preprocessing step is analogous to cleaning a crime scene before investigators begin their detailed analysis; removing irrelevant debris and preparing the scene for further investigation.

Background correction aims to eliminate the signal not originating from specific probe-target hybridization. Several methods exist:

• Subtracted background: The simplest method; subtracting the average intensity of negative controls (e.g., blank spots) from each spot’s intensity.
• Nonspecific binding correction: This method involves estimating the non-specific binding based on the intensity distribution of negative controls and subtracting it from the measurements.
• Model-based background correction: More sophisticated methods, using statistical models to account for background noise, like the ‘normexp’ method in the ‘affy’ Bioconductor package in R.

Choosing the appropriate method depends on the microarray platform and data characteristics. Often, a model-based approach is preferred for its accuracy, but requires more computational resources.

Normalization aims to remove systematic biases between arrays, making them comparable. Popular methods include:

• Quantile normalization: This method aligns the distribution of intensities across all arrays, making them have the same overall distribution. It’s a powerful method for adjusting for variations in labeling efficiency and other experimental factors.
• Robust Multichip Average (RMA): A more comprehensive method that combines background correction, normalization, and summarization steps. It’s robust against outliers and performs well with Affymetrix GeneChip data.
• LOESS normalization (Locally Weighted Scatterplot Smoothing): This technique normalizes the data by fitting a smooth curve to the relationship between the intensities of two arrays and adjusts the data based on this curve. It is suitable for two-color microarrays.

The choice of method depends on the specific microarray platform and the nature of the experimental design.

Each normalization method has strengths and weaknesses:

• Quantile normalization: Advantages: Simple, computationally efficient, and effective for removing large-scale variations. Disadvantages: Can distort biological variations if applied inappropriately.
• RMA: Advantages: Comprehensive approach, robust to outliers, handles background correction effectively. Disadvantages: Computationally intensive, more complex to implement.
• LOESS normalization: Advantages: Effective for two-color microarrays, handles dye bias effectively. Disadvantages: Can be sensitive to outliers, more complex to interpret than quantile normalization.

Selecting the optimal normalization method requires careful consideration of the data characteristics and the experimental design.

Identifying differentially expressed genes involves comparing gene expression levels between different experimental conditions (e.g., treated vs. control). This typically involves statistical tests such as:

• t-test: A common method for comparing the means of two groups. It assesses whether the difference in gene expression between groups is statistically significant.
• ANOVA (Analysis of Variance): Used when comparing more than two groups. It determines if there are significant differences in gene expression among multiple groups.
• Linear models: Offer greater flexibility for handling complex experimental designs with multiple factors.

After performing the statistical test, genes with a p-value below a chosen significance level (e.g., 0.05) and a sufficient fold-change (e.g., >2-fold) are typically considered differentially expressed. Multiple testing correction methods (like Benjamini-Hochberg) are crucial to control for false positives arising from performing many simultaneous tests.

It’s important to note that differentially expressed genes should be considered within the context of the experiment, including experimental design, sample size, and biological significance.

The false discovery rate (FDR) is a crucial concept in microarray analysis, addressing the problem of multiple hypothesis testing. When analyzing thousands of genes simultaneously, as in a microarray experiment, we’re bound to find some genes appearing differentially expressed purely by chance, even if there’s no real biological difference. The FDR helps control this. Instead of controlling the family-wise error rate (FWER), which aims to minimize the probability of making *any* false positive discoveries, the FDR focuses on controlling the expected proportion of false positives among all the discoveries. Imagine you’re sifting through sand looking for gold nuggets. FWER is like ensuring you find *no* grains of sand that you mistakenly identify as gold. FDR is more lenient, allowing for some sand (false positives) but limiting the *proportion* of sand among your ‘gold’ discoveries.

In practice, an FDR of 0.05 means that we expect 5% of the genes declared as differentially expressed to be false positives. This is a more relaxed yet still statistically rigorous approach than the FWER, especially when dealing with a large number of tests, as it allows for a larger number of statistically significant results while controlling for the overall false positive rate. This is particularly important in microarray analysis because we test thousands of genes simultaneously, and using a stringent FWER could lead to missing many truly differentially expressed genes.

Several statistical tests are commonly employed to pinpoint differentially expressed genes in microarray data. The choice often depends on the experimental design and data distribution. Here are some prominent ones:

• t-test: This is a classic method suitable for comparing the expression levels of a gene between two groups (e.g., treatment vs. control). It assesses whether the difference in means between the groups is statistically significant. Variations include the paired t-test (for paired samples) and Welch’s t-test (for unequal variances).
• ANOVA (Analysis of Variance): ANOVA extends the t-test to handle more than two groups. It determines if there are significant differences in gene expression among multiple groups.
• Linear models (e.g., LIMMA): Linear models provide a powerful and flexible framework for analyzing microarray data, accommodating complex experimental designs with multiple factors and covariates. The LIMMA (Linear Models for Microarray Data) package is a widely used and robust tool in R for this purpose.
• Non-parametric tests (e.g., Wilcoxon rank-sum test): These tests are beneficial when the data doesn’t follow a normal distribution. The Wilcoxon rank-sum test (Mann-Whitney U test) is a non-parametric alternative to the t-test.

After performing the chosen statistical test, the resulting p-values are typically adjusted for multiple testing using methods like Benjamini-Hochberg to control the FDR.

Visualizing microarray data is critical for understanding patterns and trends. Several methods effectively communicate complex datasets:

• Heatmaps: Heatmaps represent gene expression levels as a color gradient. Genes are represented as rows, and samples as columns. A red color might represent high expression, while blue represents low expression, providing a quick overview of expression patterns across genes and samples. This is particularly useful for identifying groups of genes that exhibit similar expression patterns across different conditions.
• Volcano plots: Volcano plots are excellent for visualizing differentially expressed genes. The x-axis represents the log2 fold change in expression, and the y-axis represents the negative log10 of the p-value. Genes significantly upregulated show up in the upper right quadrant, while significantly downregulated genes appear in the upper left. This visually highlights the genes that exhibit both large fold changes and statistical significance.
• Scatter plots: Scatter plots can compare the expression levels of genes between two conditions or samples. A strong positive correlation suggests that the genes are co-expressed.
• Hierarchical clustering dendrograms: Combined with heatmaps, dendrograms show the relationships between genes or samples based on their expression patterns. Similar patterns cluster together, indicating potential functional relationships or sample similarities.

Clustering in microarray analysis groups genes or samples based on their similarities in expression profiles. Think of it like sorting a collection of LEGO bricks; you’d group similar-looking bricks (e.g., color, shape) together. In microarrays, we group genes with similar expression patterns across various conditions. This can reveal co-regulated genes or functional modules involved in specific biological processes. For example, clustering might reveal a group of genes that are all upregulated in response to a particular treatment, suggesting they’re involved in a shared pathway.

Clustering helps to reduce the complexity of the data, allowing for a more manageable analysis of the large number of genes typically involved in a microarray experiment. It aids in identifying functional relationships between genes and reveals underlying biological processes.

Several clustering algorithms are used in microarray analysis, each with its strengths and weaknesses:

• Hierarchical clustering: This builds a hierarchical tree (dendrogram) representing the relationships between genes or samples. It can be agglomerative (bottom-up, starting with individual elements and merging them) or divisive (top-down, starting with the whole set and recursively splitting it). Popular linkage methods include single, complete, and average linkage.
• K-means clustering: This algorithm partitions data into k clusters, aiming to minimize the within-cluster variance. The number of clusters (k) needs to be pre-specified. This is a simpler and faster method than hierarchical clustering but requires deciding on the number of clusters beforehand.
• Self-Organizing Maps (SOMs): SOMs are neural network-based algorithms that project high-dimensional data onto a low-dimensional grid. This provides a visual representation of the data, revealing clusters and relationships between data points.

The choice of algorithm depends on the specific research question and dataset characteristics. For example, hierarchical clustering is useful for exploring the data and visualizing the relationships between data points, while K-means clustering is suitable for a more specific analysis when you know the approximate number of clusters you expect to find.

Interpreting a clustering result involves examining the resulting clusters to identify patterns and gain biological insights. This involves several steps:

• Visual Inspection: Start by visually inspecting the dendrogram (for hierarchical clustering) or the cluster assignments (for k-means). Identify clusters with clearly distinct expression patterns.
• Cluster Characterization: Analyze the genes or samples within each cluster. Determine the average expression levels of genes in each cluster and compare them across different conditions. This may involve performing functional enrichment analysis (GO analysis, pathway analysis) to determine if genes within a cluster share common biological functions or pathways.
• Biological Interpretation: Based on the characterized clusters and functional enrichment analysis, draw biological conclusions. For example, you might discover a cluster of genes upregulated in a disease state, suggesting their involvement in the disease mechanism.
• Validation: The clustering results should be validated using independent datasets or experimental techniques. This confirms the robustness and generalizability of the findings.

It’s crucial to avoid over-interpreting the results. The clusters merely suggest potential relationships; further investigation is required to confirm these relationships biologically.

Rigorous quality control (QC) is paramount to ensure reliable results in microarray analysis. Several QC checks are routinely performed:

• Background Correction: Adjusting for non-specific binding and other background signals. This improves the accuracy of gene expression measurements.
• Normalization: Correcting for systematic variations across samples or arrays, such as differences in dye labeling or RNA quality. Methods like quantile normalization or RMA (Robust Multichip Average) are commonly used.
• Outlier Detection: Identify samples or genes showing unusually high or low expression levels compared to others. These outliers can significantly impact the analysis and need investigation. Methods like boxplots and principal component analysis (PCA) can help detect outliers.
• Assessment of Data Distribution: Check if the gene expression data follows a normal distribution (or can be transformed to do so). This helps select appropriate statistical tests.
• Probe-level data analysis: If working with probe-level data instead of summarized gene expression data, it is critical to perform QC steps that check the quality of the probes and assess the reliability of the probe-level intensities. This will improve the downstream data analysis.

These QC steps ensure the data’s integrity and reliability, minimizing the influence of technical artifacts on the downstream analysis and interpretation of the results.

Missing values in microarray data are a common problem, often stemming from technical issues during the experiment. They need careful handling because ignoring them can bias downstream analyses. The best approach depends on the extent and pattern of missingness.

• Imputation methods are frequently used. These methods estimate the missing values based on the observed data. Simple methods like replacing missing values with the mean or median of the gene’s expression across all samples are quick but can obscure real biological variation. More sophisticated methods like k-Nearest Neighbors (k-NN) imputation or model-based imputation (e.g., using impute package in R) consider the expression profiles of similar samples to estimate missing values more accurately.
• Filtering is another approach, especially if the missingness is substantial or non-random. This involves removing genes or samples with a high percentage of missing values. However, this approach can lead to loss of information. A careful balance must be struck.
• Multiple imputation creates multiple plausible datasets filled in with different imputed values and then analyzes each dataset separately, combining the results at the end. This method accounts for uncertainty in the imputation process.

For example, in a study comparing gene expression in cancerous vs. healthy tissues, if a significant portion of data is missing for a specific gene in only the cancer samples, simply imputing using the mean could distort the results and mask any real difference. In such a case, either investigating the cause of missing data is crucial or applying a more advanced imputation technique would be better.

Batch effects are systematic variations in gene expression data that arise from non-biological factors introduced during the microarray experiment. These factors can include different batches of reagents, different days of processing, or even different microarray platforms. Batch effects can confound the results, masking real biological differences. They act as a form of noise, obscuring the signal.

Correcting batch effects involves adjusting the data to minimize the influence of these non-biological variations. Common methods include:

• ComBat: This popular method in the sva package in R uses an empirical Bayes approach to adjust for batch effects while preserving the biological variability. It is robust and relatively easy to implement.
• Surrogate Variable Analysis (SVA): SVA identifies surrogate variables that capture the systematic variation due to batch effects and other hidden confounders. These variables are then included in the statistical model to control for their influence.
• Removing batch-specific effects through normalization methods: Certain normalization methods, like quantile normalization, can inadvertently remove biological variation along with batch effects, which is something to be aware of. A more nuanced approach may be required.

Imagine you’re studying gene expression in different populations across various hospitals. The reagents and protocols might differ slightly at each hospital, leading to batch effects. If you don’t correct for these, you might conclude that there are genetic differences between the populations when in reality the differences are due to technical artifacts. Batch effect correction is a crucial step for accurate interpretation.

The R statistical environment, along with the Bioconductor project, is the gold standard for microarray data analysis. Bioconductor provides a rich collection of packages specifically designed for genomic data analysis.

• R provides a flexible and powerful framework for statistical computing and visualization.
• Bioconductor offers a wide range of packages for various aspects of microarray analysis, from raw data processing and normalization to statistical modeling and visualization.

Other software packages are available, such as GeneSpring or Partek Genomics Suite, but R/Bioconductor offers more flexibility and a more extensive community for support and development.

I have extensive experience using several R packages crucial for microarray analysis. Some key packages I routinely use include:

• affy: For processing Affymetrix microarray data, including background correction, normalization (e.g., RMA, MAS5), and quality control checks. I’ve used this for handling CEL files and generating expression matrices.
• limma: This package is indispensable for differential gene expression analysis, using linear models to identify genes significantly differentially expressed between different experimental conditions. It offers robust statistical methods to handle multiple comparisons and experimental design.
• edgeR: An alternative to limma, specifically designed for analyzing count data from RNA-Seq, although it has capabilities for handling microarray data. I often compare results from both limma and edgeR for robustness.
• sva: For correcting batch effects in microarray data, as I discussed earlier, using ComBat or other SVA methods.
• ggplot2: This versatile package provides elegant data visualization capabilities which I use extensively for creating publication-quality plots of gene expression profiles, heatmaps, and other relevant visualizations.

For instance, in a recent project analyzing the effects of a drug on gene expression, I used affy for preprocessing, limma for differential expression analysis, sva to control for batch effects, and ggplot2 to visualize the results, creating compelling figures demonstrating the drug’s impact on specific pathways.

Validating findings from a microarray experiment is crucial to ensure that the results are not merely artifacts. This often involves employing independent methods to confirm the microarray results.

• Quantitative PCR (qPCR): This technique can be used to validate the expression levels of a selected set of genes identified as differentially expressed in the microarray experiment. It provides a more targeted and precise measure of gene expression.
• Western blotting: This technique measures protein levels, providing another layer of validation. Changes in gene expression should ideally correlate with changes in protein levels.
• Immunohistochemistry (IHC): This technique is used to visualize protein expression in tissue samples, offering a spatial context to the gene expression data.
• Independent datasets: Validating findings on a completely independent microarray dataset from different samples or experiments strengthens the findings and eliminates experiment-specific biases. Meta-analysis techniques are useful here.

Imagine finding a gene strongly upregulated in a cancer study. Simply relying on the microarray data is insufficient. Validating the findings through qPCR, Western blotting, or comparing with an independent dataset helps establish confidence in the discovery and its biological significance. This reduces the risk of false positives.

Supervised and unsupervised learning represent different approaches to analyzing microarray data. The key distinction lies in the presence or absence of prior knowledge regarding the sample classification.

• Supervised learning uses labeled data; we know the class membership of the samples (e.g., cancer vs. healthy). The goal is to build a model that can predict the class membership of new, unseen samples based on their gene expression profiles. Methods include linear discriminant analysis (LDA), support vector machines (SVM), and various types of classification trees.
• Unsupervised learning analyzes data without prior knowledge of sample classes. The goal is to discover underlying patterns or structures in the data. Clustering algorithms like hierarchical clustering or k-means clustering are commonly used to group samples with similar gene expression profiles. Principal component analysis (PCA) is frequently used for dimensionality reduction and visualization.

In a drug efficacy study, a supervised approach might involve building a classifier to predict whether a patient will respond positively to a treatment based on their gene expression. In contrast, unsupervised learning could be used to group patients into subtypes based on their gene expression profiles, identifying potential subgroups who might respond differently to the treatment.

Despite its significant contributions, microarray technology has limitations:

• Cross-hybridization: Probes may bind to unintended sequences, leading to false positive results.
• Background noise: Non-specific binding can obscure the signal, decreasing sensitivity.
• Limited dynamic range: Microarrays may not accurately measure genes with very low or very high expression levels.
• Relative rather than absolute measurements: Microarrays primarily measure relative gene expression rather than absolute transcript quantities, making direct comparisons between different experiments or platforms challenging.
• Cost and time constraints: While costs have decreased, microarrays can still be expensive, and the experimental process can be time consuming. RNA sequencing (RNA-Seq) now offers an attractive alternative for many applications.

Understanding these limitations is crucial for proper experimental design, data interpretation, and the selection of appropriate validation methods. It’s important to compare results with other methods such as RNA-Seq to confirm findings and account for potential biases. For instance, the low dynamic range could lead to misinterpretation of gene expression changes if not properly addressed.

Microarrays and RNA-Seq are both powerful technologies used to study gene expression, but they differ significantly in their approach. Think of microarrays as a pre-printed menu with a limited selection of dishes (genes), while RNA-Seq is like ordering from a chef who can create any dish (gene) you desire, even ones not on the original menu.

Microarrays utilize pre-synthesized DNA probes attached to a solid surface. These probes hybridize with complementary cDNA from a sample, and the intensity of the signal indicates the abundance of that specific transcript. They are relatively inexpensive and established, but have limitations in sensitivity and dynamic range, and cannot detect novel transcripts.

RNA-Seq, on the other hand, directly sequences the cDNA from a sample. This provides a much more comprehensive view of gene expression, enabling the detection of novel transcripts, splice variants, and single nucleotide polymorphisms (SNPs). However, RNA-Seq is more expensive and requires more complex bioinformatics analysis.

In summary, microarrays are a well-established, cost-effective technique suitable for studying known genes, while RNA-Seq offers a more comprehensive and sensitive approach suitable for broader gene expression studies, including the discovery of novel transcripts.

My experience in designing microarray experiments encompasses all stages, from experimental design to data analysis. A well-designed experiment is crucial for obtaining reliable and meaningful results. I start by clearly defining the research question and selecting the appropriate microarray platform (e.g., Affymetrix, Agilent) based on factors like the species, the number of genes of interest, and budget constraints.

A crucial step is determining the sample size. I use power analysis to ensure sufficient statistical power to detect biologically relevant changes in gene expression. This involves considering factors such as expected effect size, variability between samples, and the desired significance level. Then, I carefully design the experimental setup, including controls (e.g., untreated samples), replicates (typically biological and technical replicates), and randomization to minimize bias. Finally, I meticulously develop a detailed protocol for sample preparation, hybridization, and scanning, ensuring consistency and minimizing variability throughout the experiment. For example, in a study comparing gene expression in two different cell lines, I would ensure that each cell line has multiple biological replicates to account for natural variations within each population, and technical replicates to account for variation within the microarray procedure itself. Careful consideration of these aspects guarantees a robust and reliable experiment, reducing noise and increasing the accuracy of results.

A microarray heatmap visually represents the expression levels of many genes across different samples. Imagine it as a color-coded spreadsheet where rows represent genes and columns represent samples. The color intensity indicates the expression level of a particular gene in a specific sample – typically, red indicates high expression, green indicates low expression, and black or yellow indicates intermediate levels.

Interpreting a heatmap involves identifying patterns and clusters. Similar expression patterns across samples suggest genes with correlated functions or regulatory mechanisms. For instance, if a cluster of genes shows high expression in a treatment group but low expression in the control group, this suggests a potential biological effect of the treatment on these genes. Hierarchical clustering, often applied to both genes and samples, helps organize the data and highlight these patterns. I also consider the scale of the heatmap; some heatmaps might be standardized to emphasize relative changes, while others display absolute expression levels. Finally, I always corroborate these visual observations with statistical analyses to confirm any suggested trends and significance.

Assessing reproducibility in microarray experiments is vital for ensuring reliability. This involves evaluating both technical and biological reproducibility. Technical reproducibility assesses the consistency of measurements within the same sample processed multiple times, highlighting the variation introduced by the microarray process itself. Biological reproducibility assesses the consistency of measurements between independent biological samples, accounting for inherent biological variability.

Several metrics quantify reproducibility. Correlation coefficients (e.g., Pearson correlation) can measure the similarity between replicates. Low correlation suggests poor reproducibility. I also use principal component analysis (PCA) to visually assess the clustering of replicates. Replicates from the same treatment group should cluster together, while distinct groups should be well-separated. Furthermore, I use statistical measures like variance analysis and coefficient of variation to quantitatively assess the variability within and between groups.

Addressing low reproducibility requires careful investigation. Possible causes include insufficient sample quality, technical errors during sample preparation or hybridization, or limitations of the microarray platform. Identifying and mitigating these issues is crucial for obtaining reliable and meaningful results.

My experience with large microarray datasets involves managing and analyzing terabytes of data generated from high-throughput experiments. This includes implementing efficient data storage and retrieval strategies, using relational databases or cloud-based storage solutions. I am proficient in using parallel computing techniques and scripting languages (e.g., R, Python) to process and analyze these datasets efficiently. This often involves breaking down the analysis into smaller, manageable tasks that can be executed in parallel, significantly reducing processing time. For example, I’ve worked with datasets containing expression data from thousands of samples and tens of thousands of genes, requiring significant computing power and optimized algorithms. Data preprocessing, normalization, and background correction are critical steps, and I utilize robust statistical methods to handle missing values and outliers.

Effective data management and storage are crucial for microarray data, which can be voluminous and complex. I have experience with various strategies, including using relational databases (e.g., MySQL, PostgreSQL) to store structured data, including experimental metadata, sample information, and normalized expression values. I also utilize specialized bioinformatics databases to integrate microarray data with other types of omics data, creating a more holistic view of biological systems. Cloud-based storage solutions (e.g., Amazon S3, Google Cloud Storage) provide scalable and cost-effective storage options for large datasets. Data security is always a primary concern, employing robust access control measures to protect sensitive information.

Furthermore, careful documentation and metadata management is crucial for long-term accessibility and reproducibility. I utilize standardized metadata formats (e.g., MIAME) to ensure data discoverability and facilitate data sharing among researchers. Data version control, using tools such as Git, ensures that changes to the data and analysis workflows are tracked and documented.

Ethical considerations in microarray data analysis are paramount. Data privacy and confidentiality are essential, especially when dealing with human samples. I strictly adhere to relevant regulations and guidelines (e.g., HIPAA, GDPR) to protect participant information. Anonymization and de-identification techniques are employed to remove any identifying information from the data, and all analyses are conducted in a secure environment.

Data integrity and transparency are also critical. I maintain detailed records of all data processing steps and analysis methods, ensuring reproducibility and allowing others to validate the results. Furthermore, I am aware of the potential biases in microarray data and employ appropriate statistical methods to account for confounding factors and minimize bias. Proper data sharing, through repositories like GEO and ArrayExpress, promotes transparency and allows other researchers to verify and build upon the findings, fostering collaboration and advancing scientific knowledge. Finally, responsible interpretation of findings is crucial, avoiding overstated claims and clearly communicating limitations of the study. Acknowledging potential limitations in the microarray methodology and interpreting results in context with other biological data ensures a responsible approach to scientific inference.