Intrinsic Limits of Read Trimming in Single-Stranded Bisulfite Sequencing

TL;DR

Single-stranded WGBS (ssWGBS) enables methylation profiling in challenging samples but introduces stochastic adaptase tailing that blurs read boundaries, creating intrinsic per-read boundary ambiguity. The authors show this boundary localization is structurally constrained and asymmetric between Read 2 (kinetic tailing) and Read 1 (geometry-driven contamination), resulting in a strictly positive Bayes error for per-read decisions. They reframe read trimming as a constrained inference problem and introduce Y-Trim, a physics-informed framework that gates trimming by observable evidence, separates Read 2 inference from Read 1 anchoring, and uses abstention and conservative projection when information is insufficient. Through controlled simulations (OOD and CpG risk) and real CCGB-34 data, the framework exposes an explicit retention–purity frontier for Read 2 while Read 1 remains information-limited, providing a principled basis for uncertainty-aware preprocessing in ssWGBS and broader texture-defined sequencing artifacts.

Abstract

Single-stranded whole-genome bisulfite sequencing (ssWGBS) enables DNA methylation profiling in low-input and highly fragmented samples, including cell-free DNA, but introduces stochastic enzymatic artifacts that complicate preprocessing and downstream interpretation. In post-bisulfite library construction, Adaptase-mediated tailing blurs the boundary between biological sequence and synthetic additions, rendering read trimming a persistent source of variability across analytical pipelines. We show that this variability reflects an intrinsic limit of per-read boundary inference rather than an algorithmic shortcoming: boundary localization is fundamentally asymmetric between paired-end reads, with Read 2 exhibiting kinetically structured artifacts that support constrained read-level inference, while apparent contamination in Read 1 arises conditionally from geometry-driven read-through events and is not well-defined at the single-read level. Even within Read 2, bisulfite-induced compositional degeneracy creates an indistinguishable regime in which genomic and synthetic origins share support under the same observable sequence evidence, implying a strictly positive Bayes error under any deterministic per-read decision rule and placing a fundamental limit on per-read boundary fidelity. By explicitly characterizing these limits, we reframe read trimming in ssWGBS as a constrained inference problem and introduce a conservative framework that operates only where supported by observable evidence (including short-range nucleotide texture), exposes interpretable trade-offs between genomic retention and residual artifact risk, and avoids forced resolution where boundaries are intrinsically unresolvable. Together, these results clarify why fixed trimming heuristics persist in practice and provide a principled foundation for uncertainty-aware preprocessing in ssWGBS.

Figures (13)

• Figure 1: Physical origin of asymmetric artifacts in ssDNA library preparation. Single-stranded DNA fragments undergo adaptor ligation followed by terminal tailing, producing a finished library molecule with two fundamentally different boundary types. The ligated adaptor constitutes a deterministic sequence template and is typically removed during standard preprocessing, whereas adaptase tailing generates a stochastic kinetic polymer with no fixed template or cut-point. Due to sequencing geometry and variable insert length, Read 2 necessarily initiates within the adaptase-tailed region, while Read 1 contains synthetic sequence only under read-through when the read length exceeds the insert length. This intrinsic asymmetry arises from sequencing geometry, and constrains boundary localization in Read 2 to a region rather than a point, while rendering read-level boundary inference for Read 1 not well-defined, motivating distinct downstream handling strategies.
• Figure 2: Read-level heterogeneity and the structural limits of fixed-length trimming in ssWGBS.(A) Base-level composition across 30 representative Read 2 sequences illustrates pronounced read-to-read variability in the transition from enzymatic tailing to genomic sequence. Guanine-rich artifacts (dark) and adenine-enriched bisulfite-converted bases (red) intermix stochastically, producing extended regions where synthetic and biological sequence cannot be reliably distinguished at the single-read level. (B) Distribution of inferred tail lengths shows a unimodal core with a pronounced long tail. A single optimal fit captures the central tendency but systematically fails to explain extreme tailing events, indicating that no unique per-read boundary can be recovered from sequence content alone. (C) Conceptual illustration of fixed-length trimming. A global cut-point inevitably induces over-trimming in reads with shorter tails, resulting in loss of genomic sequence, and under-trimming in reads with longer tails, leaving residual artifacts. Together, these panels show that constant trimming reflects a pragmatic compromise rather than a principled solution, as the true boundary is structurally heterogeneous and not directly observable from individual reads.
• Figure 3: Adaptive decision flow for read trimming with conservative fallback. Reads first pass sample-level gates to determine whether read-level inference is admissible based on sample-level conditions. For Read 2 (R2), inference separability determines whether a direct operating point can be selected along a continuous retention--purity axis. When inference is not separable, a conservative projection is applied, yielding a guarded trimming decision or abstention. Read 1 (R1) is handled independently via data-driven anchoring at the sample level. Paths leading downward indicate conservative rejection or abstention, ensuring that trimming is only applied when inference is admissible.
• Figure 4: Inference behavior and chemical consistency under controlled uncertainty. Figure 4 evaluates read-level boundary inference for Read 2 following sample-level gating under controlled synthetic regimes. The left column compares inferred trimming distributions to simulated boundary distributions across four regimes (Uniform, Bimodal, Long Tail, and Poisson), spanning idealized tails and adversarial mixtures that violate modeling assumptions. Agreement in distributional shape, rather than pointwise accuracy, indicates stable inference within regions where exact boundary recovery is intrinsically ill-defined. The right column reports per-base sequence content (PBSC) profiles of trimmed reads, confirming that trimming preserves chemical consistency without introducing systematic compositional bias. Together, these results show that the framework recovers robust structural features while remaining conservative under uncertainty.
• Figure 5: Asymmetric decision landscapes for Read 2 and Read 1 trimming. (A) For Read 2, adaptive trimming strategies collectively reveal a navigable trade-off frontier between residual synthetic artifacts and retained genomic bases. Exploratory adaptive attempts (grey) delineate the empirically admissible region, while the matrix-linear interface (purple) exposes a small number of interpretable operating points (+0, +1, +2) along this frontier. Fixed-length trimming (red) and macro-statistical trimming (dark grey) correspond to settled aggregate operating points within this space, rather than explicit exploration of the frontier itself. Retention values reflect effective genomic bases retained under admissible inference, rather than absolute biological recovery. (B) For Read 1, adaptive attempts do not form a coherent frontier and instead collapse into an unstructured family (light grey), reflecting an information-limited regime at the read level. In this setting, sample-level anchoring via a macro-statistical constant (dark grey) remains appropriate, while fixed-length trimming (red) represents a rigid but commonly used industrial fallback. Together, these panels show that adaptivity enables principled navigation of the decision space for Read 2, whereas for Read 1 the decision space itself is non-navigable, motivating fundamentally different trimming strategies for the two reads.