Stage 1: Read Structure Simulation

Segment-based Read Construction

Synthetic reads are generated as ordered concatenations of library-specific segments (e.g., random bases, ONT adapters, protocol adapters, barcodes, UMIs, cDNA, polyA/polyT). The segment order and any fixed sequence motifs are specified for each model_name input in the library definition file: seq_orders.tsv. During concatenation, the simulator emits both the nucleotide sequence and a per-base label vector identifying the originating segment.

cDNA Generation and Transcriptome Conditioning

The cDNA segment is generated either by sampling fragments from a user-provided transcriptome FASTA (--transcriptome) or, if unspecified, by generating synthetic transcripts of random bases. It is recommended to provide a transcriptome FASTA because it exposes the model to realistic intra-transcript sequence properties, including internal polyA/polyT runs and low-complexity regions. This is important because internal homopolymers can be confused with polyA/polyT tails introduced by the protocol. Training with transcriptome-derived cDNA teaches the model to resolve polyA/polyT contextually using segment boundaries rather than homopolymer presence alone.

cDNA Length as a Controlled Trade-off

cDNA length is sampled uniformly within [min_cDNA, max_cDNA]. This range is a deliberate balance:

• Shorter cDNA increases the relative proportion of fixed motifs (i.e., adapters/barcodes/UMIs), making boundary learning easier but reducing exposure to long-range sequence context.
• Longer cDNA increases length variability and generalization, but can dilute boundary signal during optimization (i.e., the model spends more capacity modeling cDNA interior), which may reduce accuracy at segment boundaries if the lengths become excessively long relative to fixed segments.

Therefore, cDNA length parameters should be tuned accordingly to match expected read structures while preserving sufficient emphasis on boundary regions.