SURF

Sequence-based sUper-Resolution Framework

Pavel Avdeyev · Kathleen Chen · Jian Zhou

UT Southwestern · University of Chicago

Explore Predictions ↓Results SURF your data →Repositories

Explore predictions

Chr8 with three signal tracks: low-coverage input, SURF prediction, and high-coverage experiment (not used for training).

Featured loci

Recover high-coverage prediction from low-coverage data

SURF turns sparse genomic measurements into higher-resolution predictions that more closely match held-out reads across coverage levels, resolutions, assays, and cell types.

0.89 vs 0.70SURF vs input correlation5% kidney DNase-seq; Pearson r with held-out reads

3×Footprint recovery gainmedian improvement over low-coverage input across DHSs

~100×Effective coverageoften comparable to much deeper data in tested regimes

Cell-type-specific signal preservationInput data vs SURF prediction

Pairwise comparison across 490 snmC-seq2 cells: sparse input correlations on the left and SURF prediction AUROC on the right — Across 490 snmC-seq2 cells, the sparse input correlation matrix (left) obscures cell-type structure, while the SURF prediction AUROC matrix (right) recovers distinct cell-type blocks in the same cell ordering.

Across assaysSpecific held-out performance

DNase-seq0.77 data-only · 0.65 sequence-onlyBoth ablations trail full SURF at 0.89 on the 5% kidney benchmark.

CAGE+30% correlation vs input0.72 SURF vs 0.55 input across five cell types on held-out chr8, chr9, and chr10.

snmC-seq2+30% single-cell correlation5 kb-binned Spearman correlation against pseudobulks from 6,983 held-out cells.

Allele-specific effects at low coverageAcross 1,387 variants, SURF reduced AF-MAE relative to the low-coverage input by ~33% at 1 read/donor and ~20% at 5 reads/donor. At a 5% false-sign rate and 2 reads/donor, SURF recovers ~68 additional known variants (of 777) over the low-coverage input.

Train SURF, calibrate with SCAN, interpret the results

Three practical paths from a first toy run to evidence-aware interpretation.

Guide

SURF your data

Train the toy model, then adapt SURF to new genomic data with the included Agent Skill.

Start the guide →Guide

SCAN predictions

Calibrate prediction intervals with the SCAN repository’s Agent Skill.

Start the guide →Guide

Interpreting results

Evaluate what SURF adds, read SCAN intervals, and decide what deserves follow-up.

Start the guide →

Repositories

Model, calibration, and analysis.

Training

jzhoulab/SURF

Toy DNase training and guided model development.

Train the reference model, then use the develop and predict Agent Skills to adapt SURF to new genomic data and export BigWig predictions.

Open on GitHub ↗Calibration

SCAN

jzhoulab/SCAN

Spline-based calibration for genomic profile predictions.

Fit calibrated intervals on held-out SURF tracks; the Agent Skill guides setup, diagnostics, and evaluation on a separate genomic region.

Open on GitHub ↗Reproducibility

SURF-manuscript

jzhoulab/SURF-manuscript

Full training code and figure notebooks for the SURF paper.

Reproduce every manuscript benchmark, analysis, and figure across DNase-seq, CAGE, and methylation.

Open on GitHub ↗

Citation

Cite the work

The manuscript is still in preparation. Please cite the GitHub repositories for now; a preprint DOI will be added when it is available.

BibTeX

@article{avdeyev_surf_inprep,
  author  = {Avdeyev, Pavel and Chen, Kathleen and Zhou, Jian},
  title   = {Sequence-based super-resolution for genomic data},
  note    = {Manuscript in preparation},
  year    = {2026}
}