Guide 03 · Interpreting results

Interpret a SURF result step by step.

A SURF track is most useful as part of a comparison. Read it beside the observed input, use SCAN to ask what is uncertain, and match the strength of the biological claim to the evidence at that locus.

1. State the question before opening the locus

Decide whether you are estimating a genomic profile, locating a footprint, comparing cell types, studying a single cell, or estimating an allele-specific effect. The same visual feature can support very different claims, and each claim needs a different control.

2. Assemble the comparison

Align these tracks to the same genome build, coordinate system, and display scale:

• Observed input: the experimental evidence available to SURF at inference time.
• SURF prediction: the enhanced profile informed by sequence and the observed input.
• SCAN intervals: uncertainty for the latent signal and for counts in a future measurement.
• Optional control: an independent replicate, deeper reference, quiet region, or cell type.

For a controlled diagnostic, run SURF prediction with --half. The repository writes the half-input, complementary target, and prediction together, so the comparison uses reads the model did not receive as its target.

3. Zoom out before judging fine structure

Begin at a scale where broad domains and peaks are visible. Ask whether SURF preserves the location and relative strength of experimentally supported events. Check a quiet region too: isolated structure in a low-signal background is easier to overinterpret than sharpening inside a supported event.

4. Ask exactly what SURF added

Zoom in only after the broad comparison makes sense. Separate three cases:

• Recovered detail: coherent fine-scale structure appears within a broader observed event.
• Stabilized signal: the prediction reduces sparse fluctuations while preserving the event.
• Model-only feature: a peak or dip appears with little local support in the observed input.

The first two can be useful candidates. Treat the third as a hypothesis that needs stronger uncertainty support, comparison across related samples, or an independent measurement.

5. Use each SCAN interval for its own question

The latent-signal interval asks where the unobserved mean genomic signal may lie. Theposterior-predictive interval asks what finite read counts a repeated experiment could produce. A latent estimate can be fairly precise while the future-count interval remains wide because read sampling adds variation. That is expected, not a contradiction.

Read interval width locally and relative to signal magnitude. A calibrated interval quantifies the uncertainty represented by SCAN; it does not remove assay bias, preprocessing errors, dataset shift, or a mistaken biological model.

Keep evaluation genuinely held out. The SCAN notebooks subtract the model input from the full coverage track before scoring because the full track contains reads the model already saw. They also cap figures at the calibration data’s upper prediction limit instead of extrapolating the dispersion spline into unsupported high-signal regions.

6. Check that biological specificity was preserved

Every cell type shares the same reference sequence, so cell-type differences must still be anchored in experimental context. Compare the locus across related cell types, donors, or cells. A convincing result preserves condition-specific peaks and relationships; a suspicious one smooths distinct samples toward the same sequence-driven profile.

7. Match the reading to the downstream task

Footprints

In the paper, footprint tracks compare signal with an expected DNase cleavage rate and then smooth the result. Because SURF preserves experimental biases such as DNase I cleavage bias, a sharp footprint is evidence to prioritize—not proof of transcription-factor occupancy by itself.

Allele-specific effects

Inspect both effect magnitude and sign. A false-sign-rate threshold controls how often selected signs disagree with the high-coverage benchmark in aggregate; it does not guarantee that one variant’s sign is correct or establish the mechanism of the effect.

Single-cell profiles

Ask whether the gain comes with preserved cell-type specificity and cell-to-cell variation. A cleaner profile is not useful if biologically meaningful differences between cells have been erased.

8. Classify the finding

Supported candidate

The observed input supports the broad event, SURF adds coherent detail, SCAN uncertainty is usable for the intended claim, and the feature is consistent across relevant samples or controls.

Needs follow-up

The feature is biologically interesting but weak in the input, uncertain under SCAN, sensitive to track scaling, or inconsistent across related samples. Keep it, but label the missing evidence.

Do not interpret alone

The feature appears only after prediction, lies in a bias-prone or poorly covered region, has unsuitable uncertainty, or depends on a preprocessing or coordinate mismatch.

9. Record the reasoning, not just the screenshot

Biological question:
Region and genome build:
Observed support:
Structure added by SURF:
SCAN interval used and why:
Cell-type / replicate comparison:
Known assay or preprocessing bias:
Claim supported at this locus:
Follow-up needed:

Keep these limits in view

• SURF preserves experimental biases such as DNase I cleavage bias. Downstream analyses should treat them the same way they would for the raw experiment.
• SURF estimates are high-confidence predictions, not experimental measurements. The framework improves over the input on average but is not guaranteed correct at every locus.
• Best used to generate hypotheses and guide follow-up experiments, for example targeted sequencing of a region of interest.

10. Turn uncertainty into the next experiment

Use SURF to rank hypotheses, not merely to make tracks look smoother. Prioritize regions where added structure changes a biological decision, then choose the follow-up that addresses the weakest evidence: deeper sequencing for sampling uncertainty, another assay for mechanism, or independent samples for reproducibility.