Graph Neural Network Variant Caller

The foundational data structure where nodes represent individual sequencing reads and the reference genome, while edges encode alignment overlaps and candidate allele relationships. Unlike a rigid pileup, this graph captures the full topology of read connectivity, preserving information about paired-end insert sizes and split-read mappings that are critical for resolving structural variants and indels. The graph is constructed by chaining alignments, creating a rich relational structure that explicitly models the evidence for each allele.

The core learning mechanism where nodes iteratively exchange and aggregate feature vectors with their neighbors. Each message function computes information to send along edges, while an update function (typically a small neural network) integrates received messages to refine each node's hidden state. After multiple rounds of propagation, the read nodes contain contextualized embeddings that reflect not just their own sequence but the consensus of all overlapping reads, enabling the model to distinguish true variants from sequencing errors.

A preprocessing stage that scans the alignment graph to identify positions with evidence for variation. This module generates a set of candidate alleles—including SNPs, insertions, deletions, and complex substitutions—by analyzing read pileups, soft-clipped bases, and discordant read pairs. By proposing candidates before the neural network runs, the system avoids the computational cost of evaluating every possible genotype at every position, focusing the message-passing computation only on plausible variant sites.

The final layer that transforms the refined node embeddings into calibrated genotype probabilities. For each candidate variant, the decoder aggregates the hidden states of all reads and reference nodes in the local subgraph and outputs a diploid genotype likelihood across all possible allele combinations (homozygous reference, heterozygous, homozygous alternate). The decoder is trained to produce well-calibrated probabilities, enabling direct integration with downstream Variant Quality Score Recalibration pipelines.

A critical design choice where the properties of each alignment are encoded as edge features in the graph. These features include mapping quality, base quality scores, CIGAR string operations, strand orientation, and edit distance. By providing the neural network with rich information about the quality and nature of each read-reference relationship, the model can learn to downweight unreliable alignments and upweight high-confidence evidence during message passing.

A hierarchical aggregation strategy that operates at multiple resolutions within the graph. Local pooling aggregates information within a single candidate variant's neighborhood, while global pooling captures broader genomic context such as coverage depth trends and regional mappability. This multi-scale approach allows the model to condition its variant calls on both local read evidence and regional characteristics, reducing false positives in low-complexity or repetitive regions.

The statistical foundation that graph neural network callers refine through message passing. Traditional likelihoods compute:

P(Read Data | Genotype = A/B)

Graph neural networks enhance this by propagating information along edges, allowing a read's evidence to influence probability estimates at connected loci. This is critical for:

• Resolving low-frequency variants where individual read support is sparse
• Disambiguating heterozygous calls in regions with systematic sequencing errors
• Jointly modeling variants that are physically linked on the same haplotype

The process of assigning variants to maternal and paternal chromosomes, which graph neural network architectures naturally support through their message-passing topology. Key methods include:

• Read-backed phasing: Using paired-end reads spanning multiple heterozygous sites to link alleles
• Population-based phasing: Leveraging linkage disequilibrium patterns from reference panels

Graph neural networks can perform joint variant calling and phasing by allowing allele assignments to propagate along edges representing physical read connections, improving both accuracy and haplotype resolution.

A post-calling machine learning step that assigns calibrated error probabilities to variant calls. While GATK's VQSR uses a Gaussian mixture model on hand-engineered annotations, graph neural network callers can integrate this directly:

• Learned embeddings from the graph serve as rich, automatically derived features
• The final classification layer outputs well-calibrated probabilities without a separate recalibration step
• Reduces the need for external truth sets by learning error modes from the data distribution itself