Pileup Image Encoding

The core innovation is mapping biological data to standard image channels. Each pixel column represents a genomic position, and each row is a read. The red channel encodes the reference base and reads matching it. The green channel encodes reads with the alternate allele. The blue channel encodes base quality scores. An additional alpha channel often represents strand orientation (forward vs. reverse). This allows a CNN to learn spatial features like strand bias artifacts directly from the pixel patterns.

Encoding does not occur genome-wide. A candidate variant locus is first identified by a traditional caller or heuristic. A fixed-size window (e.g., 221x100 pixels for DeepVariant) is centered on the candidate site. This window captures the local alignment context, including flanking reference sequence and nearby reads that may span the variant. The fixed dimensions ensure a consistent input tensor size for the neural network, regardless of the actual depth of coverage at the locus.

Before pixel generation, reads overlapping the candidate window are locally realigned to the reference using algorithms like Smith-Waterman. This corrects for alignment artifacts near indels. Reads are then sorted by haplotype or alignment score to create a consistent vertical ordering. This sorting is critical: it transforms an unordered set of reads into a spatially coherent image where the CNN can learn vertical patterns associated with heterozygous alleles or mapping errors.

Base quality scores (Phred scale) are not discarded but embedded directly into pixel intensity. A high-quality base matching the reference produces a bright pixel in the red channel. A low-quality base produces a dim pixel. This allows the CNN to learn to down-weight noisy signals automatically. The model learns to distinguish a true variant supported by high-quality mismatches from a sequencing error supported by low-quality mismatches, all from the visual pattern of pixel brightness.

To detect strand bias artifacts, the encoding separates reads by strand. Forward-strand reads and reverse-strand reads are often placed in distinct rows or encoded with different alpha values. A true heterozygous variant should show the alternate allele on both strands. If the green channel signal appears only in the forward-strand rows, the CNN learns to recognize this as a systematic artifact, likely from oxidative damage during library preparation, and will down-weight the variant call.

The reference genome sequence is always included as the top row of the pileup image, providing a visual anchor for the CNN. This row is encoded in the red channel. The model learns to compare each read row against this reference row to identify mismatches. This explicit representation of the reference eliminates the need for the model to infer the expected base from the read pileup, simplifying the learning task and improving accuracy at repetitive loci.