FASTQ Simulation

Accurate simulation requires modeling the sequencing error profile of specific platforms. This includes substitution errors (mismatches) and indel errors (insertions/deletions), which vary systematically based on sequence context, base position within the read, and the specific chemistry of the sequencing instrument. A high-fidelity simulator replicates these biases to ensure downstream variant callers are tested against realistic noise.

A defining feature of FASTQ is the encoding of a base quality score (Phred score) for every nucleotide. A simulator must generate these scores realistically, modeling the characteristic decline in quality toward the 3' end of a read. The distribution of quality scores must match empirical data from real sequencers, as tools like Trimmomatic and BWA-MEM use these scores for read filtering and mapping confidence.

Simulators must model the stochastic nature of shotgun sequencing. Key parameters include:

• Mean coverage depth: The average number of reads covering a base (e.g., 30x for whole-genome).
• GC content bias: Systematic under- or over-representation of reads in regions with extreme guanine-cytosine content.
• Coverage uniformity: Real data shows Poisson-like or over-dispersed distributions, not perfectly uniform coverage.

To benchmark variant callers, simulators must inject known genetic variants at precise variant allele frequencies (VAF). This allows developers to measure sensitivity and precision at specific VAF thresholds. For somatic variant detection, the simulator must model tumor purity and clonal heterogeneity, generating reads with a mixture of reference and alternate alleles at a defined ratio.

Modern sequencing uses paired-end reads, where both ends of a DNA fragment are sequenced. A simulator must generate read pairs with a realistic insert size distribution (the distance between the paired reads). This distribution is typically modeled as a normal or log-normal curve. Correct insert size modeling is critical for testing structural variant detection and read alignment algorithms.

Real sequencing data contains technical artifacts that pipelines must handle. A comprehensive simulator generates:

• Adapter contamination: Residual sequencing adapter sequences ligated to read ends.
• Optical duplicates: Reads arising from the same cluster on a flow cell, identified by identical mapping coordinates.
• Chimeric reads: Artifactual reads formed by the ligation of two unrelated DNA fragments.

The core process of creating artificial short DNA fragments that mimic the output of high-throughput sequencing machines. This involves modeling realistic error profiles and base quality scores to ensure downstream bioinformatics tools cannot distinguish synthetic reads from real ones. Key components include:

• Simulation of substitution errors and indel errors
• Modeling of GC content bias across read positions
• Generation of paired-end read pairs with realistic insert size distributions

A per-nucleotide confidence metric encoded in the FASTQ format using Phred quality scores. Each base is assigned a score Q, where the error probability P = 10^(-Q/10). A score of 30 (Q30) indicates a 0.1% error probability. Generative models must realistically simulate the declining quality toward read ends characteristic of sequencing-by-synthesis chemistry. Quality scores are encoded as ASCII characters in FASTQ files, with the offset (+33 for Sanger format) being a critical simulation parameter.

A statistical model of systematic and random errors introduced during DNA sequencing. Realistic simulation requires modeling:

• Phasing errors: Signal decay as synthesis cycles progress
• Pre-phasing errors: Residual signal from previous cycles
• Tiling artifacts: Position-specific biases on flow cells
• Base-calling errors: Misclassification of fluorescent signals These profiles are often derived empirically from real sequencer runs and applied as conditional probability matrices during simulation.

A specialized Generative Adversarial Network framework designed to generate discrete nucleotide sequences. Unlike standard GANs that operate on continuous data, SeqGAN uses reinforcement learning-based policy gradients to overcome the non-differentiability of discrete token generation. The discriminator evaluates complete sequences, providing reward signals that guide the generator toward producing biologically plausible reads with realistic k-mer frequency distributions and motif preservation.

A metric for evaluating synthetic genomic data quality by comparing the distribution of generated sequences to real sequences in a learned feature space. Analogous to the Frechet Inception Distance (FID) in computer vision, it computes the Wasserstein-2 distance between multivariate Gaussian distributions fitted to feature embeddings. Lower scores indicate higher fidelity. This metric is sensitive to both mode collapse and mode dropping in generative models.

An evaluation paradigm where a predictive model is trained exclusively on synthetic genomic data and tested on real data. The performance gap compared to a model trained on real data quantifies the utility of the synthetic data. A minimal gap indicates the synthetic data captures the relevant biological signal. This approach is preferred over distributional metrics when the end goal is a specific downstream task like variant calling or gene expression prediction.