CANDI

Epigenome imputation began as a way to predict experiments nobody had run, but researchers quickly noticed something more useful: imputed tracks are often cleaner than the measurements themselves, because a model that integrates patterns across assays, cell types and loci averages away batch effects and technical noise. Denoising by re-imputation is now routine, and it exposes two weaknesses in existing tools. Every prior method operates on idealized processed signal such as fold enrichment over control, assuming processing has already removed batch effects — an assumption that undermined the ENCODE Imputation Challenge, where a naive baseline beat most entrants on subtle train/test differences. And denoising a specific experiment normally means retraining the model without it.

CANDI, the Confidence-Aware Neural Denoising Imputer, comes from Mehdi Foroozandeh, Abdul Rahman Diab and Maxwell Libbrecht in the School of Computing Science at Simon Fraser University. It changes three things at once. It predicts raw read counts and takes experimental covariates — sequencing depth, read length, coverage, run type — as explicit inputs rather than pretending they have been normalized away. It can optionally accept a low-quality existing measurement of the target assay and use it while predicting, so denoising requires no retraining. And it emits probability distributions instead of point estimates, giving every position a calibrated uncertainty.

#Key Features

• Raw counts with covariates: Inputs and outputs are read counts alongside the four experimental covariates, so the model learns the relationship between sequencing depth and observed signal instead of inheriting it from a preprocessing step.
• Dual output distributions: Each assay and position gets a negative binomial over raw counts and a Gaussian over processed signal in log Poisson p-value units, so both the count and signal views come with error bars.
• Zero-shot across cell types: At test time the model sees only DNA sequence and the observed assays of the target sample — no cell type embedding and no positional metadata — forcing it to generalize rather than memorize.
• Denoising without retraining: A degraded existing experiment can be supplied as input when predicting that same assay, which is what makes practical denoising possible on a fixed checkpoint.
• Assay prioritization: Because uncertainty is quantified, the model can be queried for which additional assay would most improve prediction of a target in a given cell type.

#Technical Details

For each 30 kb locus at 25 bp resolution — 1,200 genomic positions — the encoder runs parallel Conv1D towers over the epigenomic count matrix and the one-hot DNA sequence, downsampling both to a shared 150-position context, fuses in embedded covariates, and passes the result through four transformer encoder layers with 9 attention heads and relative positional encoding. Two deconvolutional decoder towers then restore resolution and feed negative binomial and Gaussian distribution layers. Training is self-supervised with two objectives: masking entire assays and reconstructing them, and downsampling reads to simulate low-depth experiments and recovering the original. Training used 3,000 non-overlapping 30 kb regions containing candidate cis-regulatory elements — 90 million base pairs, about 3% of the human genome — with chromosome 21 excluded and reserved for testing. Two datasets back the evaluation: the ENCODE Imputation Challenge set of 35 assays across 50 biosamples, and an extended collection of 3,064 ENCODE biosamples merged by cell type into 361 cell types. On EIC, CANDI matches Avocado and the challenge winners on Spearman correlation and MSE despite using no cell type metadata; on the extended set Pearson correlations exceed 0.8 for active marks such as H3K27ac and H3K4me3. Confidence intervals are near-perfectly calibrated above the 90% level and conservative below it, partly because the predicted negative binomial median exactly matches the observed count 38% of the time.

#Applications

The target user is anyone who runs downstream analysis on ENCODE-style tracks — chromatin state annotation, regulatory element calling, expression modelling — and who has to decide whether a noisy or shallow experiment is trustworthy. CANDI denoises that experiment in place and attaches a confidence estimate, so high-signal, high-confidence regions separate from high-signal, low-confidence ones. The paper shows this matters: confidently predicted H3K4me3 peaks overlap transcription start sites far more often than low-confidence ones. The same estimates support experiment design, ranking which assay to run next.

#Impact

CANDI is a bioRxiv preprint awaiting peer review. Its contribution is less a leaderboard result — imputation accuracy is comparable to the challenge winners rather than ahead of them — than a reframing of what an imputation model should output, with calibrated uncertainty and raw counts replacing point estimates on processed signal. The repository provides training, inference and evaluation code plus the full ENCODE data pipeline, but no pretrained checkpoint is distributed and the repository carries no license file, so running the model currently means training it. The preprint is released under CC-BY-ND.