This plot shows how many false positives you can expect at different confidence cutoffs. It helps you choose a threshold that balances discovering real interactions against accepting some false positives.

The x-axis is the posterior probability threshold; the y-axis is the estimated FDR. The curve is derived from simulation-based calibration. A typical target is FDR below 5% (dashed line).

This plot shows what fraction of true interactors you would recover at each confidence threshold. Stricter thresholds miss more real interactions but produce fewer false positives.

The x-axis is the posterior probability cutoff. The y-axis is sensitivity (true positive rate). The curve is estimated from synthetic data generated to match your experiment's characteristics.

This plot evaluates the overall ability of the analysis to distinguish true interactors from background proteins. A curve hugging the top-left corner indicates excellent discrimination.

The ROC curve plots sensitivity (true positive rate) against 1 minus specificity (false positive rate) at all possible thresholds. The area under the curve (AUC) summarizes performance: 1.0 is perfect, 0.5 is random.

This plot checks whether the reported confidence values are trustworthy. If a protein is assigned 80% probability, it should truly be an interactor about 80% of the time.

Points on the diagonal line indicate well-calibrated probabilities. Points above the diagonal mean the analysis is underconfident; below means overconfident. Bars show how many proteins fall in each confidence bin.

This table suggests optimal confidence thresholds for different analysis goals, such as maximizing the number of discoveries or minimizing false positives.

Each row shows a threshold strategy with its estimated FDR, sensitivity, specificity, and the number of proteins that would be called significant. Choose based on whether you prioritize completeness or precision.

This plot compares confidence values before and after calibration. Calibrated probabilities should align more closely with the diagonal, meaning they better reflect actual interaction rates.

Blue points show raw posterior probability values; orange points show calibrated values. If calibration improved the estimates, orange points will be closer to the diagonal than blue.

This plot shows the mathematical transformation applied to adjust raw confidence values into calibrated probabilities.

The x-axis is the raw posterior probability; the y-axis is the calibrated value. The calibration function (Platt scaling) is a logistic regression fitted on simulation ground truth. Points near the diagonal indicate minimal adjustment was needed.

This table reports how well the calibration holds up when tested on data not used to fit the calibration model, ensuring the correction generalises beyond the training data.

ECE (Expected Calibration Error) measures the average gap between predicted probabilities and observed frequencies. Lower ECE means better-calibrated posteriors. Values are from cross-validation folds.

This plot checks whether the declared false discovery rates match the actual error rates observed in simulations. Points near the diagonal mean the FDR estimates are accurate.

The x-axis is the declared FDR (what you are told); the y-axis is the actual FDR (measured from simulation ground truth). A well-calibrated analysis produces points along the diagonal.

This plot summarises how similar the protein rankings are across all tested prior specifications. High correlations mean the ordering of interactors is stable regardless of prior choice.

Each point is the Spearman rank correlation between the default analysis and one alternative prior sensitivity specification. Correlations above 0.95 indicate excellent robustness.

This matrix shows how similar the protein rankings are between every pair of prior specifications. All-blue means results are completely insensitive to prior choice.

Each cell is the Spearman rank correlation between two prior specifications. Values close to 1.0 indicate that the protein ranking is preserved. Off-diagonal entries below 0.95 flag settings where the ranking changes meaningfully.

Each whisker shows the full range of posterior probability for one protein across all tested priors. Proteins whose whiskers cross the red dashed line (P=0.5) change classification depending on prior choice -- these are flagged in red.

Proteins are sorted left-to-right by decreasing sensitivity range. The dashed red line marks the P=0.5 decision boundary. Red markers indicate proteins that cross this boundary across the prior grid. Narrow whiskers indicate robust classification.

For the most sensitive proteins, this plot shows the exact posterior probability under each prior setting as individual points. Tight clusters mean the result is stable; spread-out points reveal sensitivity to prior choice.

Each column is one protein (the top 15 most sensitive). Each dot within a column represents the posterior probability under one prior setting. The red dashed line marks P=0.5. Proteins with points on both sides of this line change classification across priors.

This chart shows how the relative importance of the two statistical models (Copula and 3c-EM) changes across different prior specifications. Stable weights indicate that model selection is robust to prior choice.

Each bar shows the LOO (leave-one-out) stacking weight for the 3c-EM and Copula models at each BMA prior grid point. Weights sum to 1.0. Large shifts across the grid indicate sensitivity of model selection to prior specification.

This plot shows how the statistical model has classified each protein into one of three groups: background (non-interactors), ambiguous, or genuine interactors. Clear separation between groups indicates the model can reliably distinguish interactors from noise.

Each point is a protein, positioned by its enrichment Bayes factor (x) and correlation Bayes factor (y) on log10 scale. Colour indicates component assignment from the 3c-EM model.

These plots show how well the statistical model captures the distribution of evidence scores in your experiment. Good fits mean the model assumptions match your data.

Each panel shows the observed distribution of a Bayes factor type (histogram) overlaid with the fitted mixture components. The background (H0) component captures the null distribution; the H1 component captures genuine interactors.

This matrix provides a quick pass/fail overview of key quality checks on the statistical model. Green cells indicate the model is behaving as expected; yellow or red cells flag potential issues.

Each cell tests a specific aspect of model fit: marginal goodness-of-fit (KS test for enrichment/correlation, chi-squared test for detection), component separation, contamination between H0 and H1, and within-class independence. Click individual cells for details.

This plot compares the mass spectrometry evidence for each interaction with the structural prediction confidence from AlphaFold. Proteins in the upper-right have both strong MS evidence and plausible predicted structures.

The x-axis shows -log₁₀(PEP) from the MS-only analysis (higher values = more confident); the y-axis shows the ipTM score from AlphaFold structure prediction. An ipTM above 0.6 suggests a plausible physical interaction; below 0.2 provides no structural support.

This plot shows how structural evidence from AlphaFold changes interaction error rates. Proteins that shift downward gained support from structural predictions (lower PEP = more confident). The log₁₀ scale reveals changes among high-confidence interactions that would be compressed near zero on a linear scale.

Both axes show -log₁₀(PEP) (higher = more confident). The x-axis is the MS-only evidence; the y-axis is the updated value after incorporating structural Bayes factor evidence. Points above the diagonal gained confidence from docking; below lost confidence.

This table lists all proteins for which structural predictions were obtained, with their structural quality scores and the resulting update to interaction probabilities.

Key columns: ipTM (interface quality), pLDDT (local structure confidence), pDockQ (docking quality), and the structural Bayes factor used to update the MS-based posterior. Values are from AlphaFold Server predictions.

BayesInteractomics expects intensity values on a log2 scale. If your data is on the original linear scale (raw intensities), the statistical models may produce unreliable results due to the highly skewed distribution of untransformed protein abundances.

Scale detection checks the maximum observed intensity value. Log2-transformed AP-MS data typically falls in the range of 10-35. Maximum values above 1000 strongly suggest linear-scale (untransformed) data. If flagged, apply log2 transformation before re-running the analysis.

Replicate correlation measures how consistently each replicate reproduces the same protein abundance pattern. In AP-MS experiments, Spearman correlations above 0.80 between replicates indicate good technical reproducibility.

Each cell shows the pairwise Spearman rank correlation between two replicates within the same group (sample or control). Green cells indicate strong agreement (r >= 0.80), yellow indicates moderate agreement (0.60 <= r < 0.80), and red indicates poor agreement (r < 0.60). The number of shared non-missing proteins is shown on hover.

Missing values are common in AP-MS data because low-abundance proteins may not be consistently detected across replicates. However, if one replicate has substantially more missing values than others, it may indicate a failed injection or degraded sample.

Each bar shows the fraction of proteins with missing intensity values for one replicate. Bars are grouped by experimental group (sample/control). A replicate is flagged if its missing fraction is more than 2x the group median. Green indicates normal missingness, yellow indicates elevated missingness (2-3x median), and red indicates extreme missingness (>3x median).

The shape of the intensity distribution reveals technical artifacts. Well-behaved AP-MS data on a log2 scale typically shows an approximately normal distribution. Bimodal distributions, heavy tails, or value spikes suggest instrument issues or data processing artifacts.

This section summarizes the distribution shape checks across all replicates. High excess kurtosis (>7) indicates heavy tails; negative kurtosis (<-1.2) suggests bimodality; spike fraction >10% indicates value pile-up at a boundary. These checks do not require corrective action but help contextualize downstream results.

Principal Component Analysis (PCA) reduces the full protein abundance matrix to its most important axes of variation. If the experiment worked well, the biggest source of variation (PC1) should separate your bait samples from your controls.

Each point represents one replicate. The x-axis (PC1) and y-axis (PC2) show the two largest axes of variation, with the percentage of total variance they explain. Sample points are colored separately from control points. If samples and controls overlap heavily on PC1, the experimental condition may not be the dominant effect, which could reduce the sensitivity of the Bayesian analysis.

PCA on the post-imputation log-intensity matrix. Use the color dropdown to switch among condition / replicate / experiment / protocol annotations.

SVD-based PCA on the (n_samples × n_proteins) matrix after cascading complete-case filtering. PC1 and PC2 percent variance is shown in the axis labels.

Non-linear sample embedding via UMAP. Preserves both local clusters and global structure better than PCA on heavy-tailed data.

UMAP.jl with seed = , n_neighbors clamped to min(15, n_samples − 1).

UMAP on the 5-feature posterior vector per protein. Points are coloured by classification (H0 / Agnostic / H1).

Features: [log10(bf_enrichment), log10(bf_correlation), log10(bf_detected), posterior_prob, log2FC_mean] with the clamp log10(max(bf, 1e-12)), z-scored before UMAP.

This table lists every prior distribution used in the analysis. All priors are weakly informative unless stated otherwise. Prior sensitivity was assessed by varying key parameters; see the Sensitivity tab for results. Following BARG guidelines (Kruschke, 2021), each prior is justified in the rightmost column.

The detection Bayes factor BF₁₀ compares two hypotheses: H₁ (detection rate differs between sample and control) versus H₀ (detection rates are equal). The Beta(, ) prior is symmetric and weakly informative, centred at 0.5, allowing the data to drive inference. A BF₁₀ > 3 indicates moderate evidence for differential detection (Kass & Raftery, 1995).

Intensity means are assigned Normal(mu=, sigma=) priors, and precisions receive Gamma(shape=, scale=) priors. The enrichment BF₁₀ tests H₁ (log₂FC ≠ 0) against H₀ (log₂FC = 0). Posterior medians with 95% highest-density intervals (HDI) are reported for effect sizes.

The slope prior is . The correlation BF₁₀ compares H₁ (|slope| > ) against H₀ (|slope| ≤ ).

Platt scaling fits a logistic transformation: P_calibrated = logistic(a · logit(P_raw) + b), where a and b are estimated from simulation ground truth. The ECE safety guard prevents calibration from degrading probability estimates. Synthetic data are generated using fitted model parameters from the real data, preserving realistic correlation structures.

This table lists all the settings used to run your analysis, ensuring full reproducibility. If you share these parameters, another researcher can replicate your exact analysis.

Parameters include the number of control and sample replicates, the reference protein index, prior sensitivity ranges, evidence combination method (BMA, copula, or 3c-EM), and all model-specific settings.

This code block provides the exact Julia commands needed to reproduce this analysis, including all parameter settings and package versions.

Copy this code block into a Julia session with BayesInteractomics installed to regenerate this report. The package version and all configuration values are captured for exact reproducibility.