Epigenomic Uncertainty Quantification

Captures model ignorance—the uncertainty arising from incomplete knowledge of the optimal parameters or architecture. This uncertainty is reducible with more training data or better model design.

• Quantified via Bayesian neural networks that place distributions over weights rather than point estimates
• Measured by the variance across an ensemble of models trained with different initializations
• High epistemic uncertainty signals regions where the model has insufficient training support, such as rare regulatory elements in uncharacterized cell types
• Directly informs active learning strategies to prioritize experimental validation of the most uncertain predictions

Represents inherent data noise—the irreducible stochasticity in biological systems and measurement processes. This uncertainty persists even with infinite training data.

• Arises from biological variability between cells, stochastic chromatin remodeling, and technical noise in sequencing assays
• Modeled by predicting a variance parameter alongside the mean prediction, enabling heteroscedastic output distributions
• Critical for distinguishing true biological heterogeneity from measurement artifacts in single-cell epigenomic data
• High aleatoric regions often correspond to poised or bivalent chromatin states where regulatory activity is intrinsically variable

A practical approximation to Bayesian inference that applies dropout at inference time to generate multiple stochastic forward passes, producing a predictive distribution without architectural changes.

• Each forward pass randomly drops different neurons, creating an implicit ensemble from a single trained model
• The mean of multiple passes provides the prediction; the variance quantifies epistemic uncertainty
• Widely adopted in genomic models like Basset and DeepSEA due to minimal computational overhead
• Particularly effective for identifying out-of-distribution sequences where the model extrapolates beyond its training manifold

Trains multiple independent models with different random initializations and data orderings, then aggregates their predictions to produce well-calibrated uncertainty estimates.

• Captures functional diversity in the loss landscape, providing more robust uncertainty than single-model methods
• Each ensemble member converges to a different local minimum, exploring distinct but equally plausible solutions
• The spread of predictions across ensemble members directly quantifies epistemic uncertainty at each genomic locus
• Outperforms Monte Carlo dropout for sequence-to-epigenome models like Enformer when computational budget permits

Evaluates whether predicted confidence scores align with empirical accuracy—a well-calibrated model should be correct 90% of the time when it assigns 90% confidence.

• Measured via Expected Calibration Error (ECE) which bins predictions by confidence and compares to observed accuracy
• Genomic models often exhibit overconfidence on regulatory variants in underrepresented genomic contexts
• Temperature scaling—a post-hoc method that learns a single parameter to soften output probabilities—improves calibration without affecting ranking performance
• Critical for clinical applications where miscalibrated uncertainty can lead to overconfident variant pathogenicity predictions

Integrates uncertainty quantification into in-silico mutagenesis workflows to distinguish high-confidence regulatory variants from ambiguous predictions requiring experimental follow-up.

• For each virtual mutation, the model reports both a predicted effect size and an uncertainty interval
• Variants with high effect size and low uncertainty are prioritized for functional validation
• Variants with high uncertainty—regardless of predicted effect—are flagged for additional training data collection in that genomic context
• Enables risk-stratified reporting in clinical genomics, where false positives carry significant consequences

The foundational distinction in uncertainty quantification. Epistemic uncertainty is model ignorance—reducible by collecting more training data from underrepresented genomic regions. Aleatoric uncertainty is inherent noise in the data itself, such as stochastic chromatin accessibility or sequencing depth variability. A well-calibrated model outputs high epistemic uncertainty for a novel promoter sequence never seen during training, but high aleatoric uncertainty for a genomic region with intrinsically noisy histone modification signals across replicates.

A practical Bayesian approximation technique that enables uncertainty estimation without modifying existing architectures. By applying dropout at inference time and running multiple stochastic forward passes, the variance across predictions quantifies epistemic uncertainty.

• Implementation: Keep dropout layers active during inference on a trained Basset or DeepSEA model
• Output: Mean prediction and variance for each epigenomic track at each genomic bin
• Use case: Flagging variant effect predictions where the model's confidence is low due to rare sequence contexts

Training multiple independent models with different random initializations and evaluating their prediction spread provides a robust measure of epistemic uncertainty. Unlike single-model methods, ensembles capture uncertainty arising from the loss landscape's geometry.

• Architecture: 5-10 independently trained Enformer or Basenji2 instances
• Aggregation: Mean prediction across ensemble members; standard deviation as uncertainty
• Advantage: Better calibrated than Monte Carlo dropout for out-of-distribution sequences from unseen cell types

A distribution-free framework that provides statistically rigorous prediction intervals with guaranteed coverage probabilities. For epigenomic regression tasks, conformal prediction outputs a band around the predicted chromatin accessibility or methylation level that contains the true value with a user-specified confidence level (e.g., 95%).

• Calibration set: Held-out genomic regions not used in training
• Output: Prediction intervals that adapt to local sequence complexity
• Regulatory application: Identifying variants where predicted effect sizes fall within the conformal interval, indicating insufficient evidence for functional impact

Instead of learning point estimates of weights, Bayesian neural networks learn probability distributions over model parameters. This provides a principled framework for uncertainty quantification in epigenomic predictions.

• Variational inference: Approximates the true posterior distribution over weights
• Last-layer Bayes: A computationally efficient compromise where only the final layer is Bayesian
• Output: Full predictive distribution for each epigenomic track, enabling credible intervals for enhancer activity predictions

Quantifying how well a model's confidence aligns with its accuracy. A perfectly calibrated model predicting chromatin accessibility with 80% confidence should be correct 80% of the time.

• Expected Calibration Error (ECE): Measures the weighted average gap between confidence and accuracy across bins
• Reliability diagrams: Visual plots of confidence vs. observed accuracy
• Brier score: Mean squared error between predicted probabilities and binary outcomes for peak calling tasks
• Temperature scaling: A post-hoc method that adjusts model confidence without affecting accuracy rankings