Paired-End Read Mapping

• Paired-end read mapping is the process of aligning two short DNA fragments from opposite ends of genomic segments, leveraging insert-size constraints for enhanced precision.
• It employs a joint seed-based filtering algorithm and lightweight bitwise alignment to significantly reduce computational load compared to traditional dynamic programming methods.
• The approach integrates specialized hardware acceleration to achieve high-throughput, energy-efficient genome analysis while maintaining robust variant calling accuracy.

Paired-end read mapping refers to the process of aligning pairs of short DNA fragments, sequenced from both ends of longer genomic segments, to a reference genome. This approach is favored in modern genome analysis for its higher accuracy and ability to support advanced inference tasks. Mapping paired-end reads is computationally intensive due to the need to evaluate possible placements for both reads while respecting their expected genomic proximity (the insert-size window). Recent developments have emphasized joint filtering algorithms and hardware-algorithm codesign, exemplified by GenPairX, a system that implements an efficient pipeline combining seed-based filtering, lightweight alignment, and specialized accelerator architecture for throughput and energy efficiency (Eudine et al., 27 Jan 2026).

1. Joint Paired-End Filtering Algorithm

The GenPair filter exploits the requirement that both ends of a paired-end read map within a predefined distance ($\Delta$) in the genome. For each read pair $(R_1, R_2)$, GenPair extracts $k$-mer seeds $S_1 = \{s_1, s_2, s_3\}$ (from $R_1$) and $S_2 = \{t_1, t_2, t_3\}$ (from $R_2$), typically three nonoverlapping 50 bp seeds per read. A hash-based index called SeedMap maps each seed to all genome locations that match exactly. The lists $L_{1i}$ and $L_{2j}$ for seeds from both reads are merged, yielding $L_1$ and $(R_1, R_2)$0.

Candidate mapping pairs are defined as:

$(R_1, R_2)$1

Only pairs in $(R_1, R_2)$2 proceed to alignment; all others are pruned, substantially reducing the computational load.

The filtering ratio is:

$(R_1, R_2)$3

On human-genome short-read data, GenPairX achieves $(R_1, R_2)$4, whereas single-read filters achieve less than 40% filtration on paired-end data.

The filtering step is realized by the following pseudocode:

$S_2 = \{t_1, t_2, t_3\}$2 Complexity is $(R_1, R_2)$5 per read-pair.

Hash-index false positives are suppressed with a 32-bit xxHash ($(R_1, R_2)$6 per seed). The distance threshold $(R_1, R_2)$7 is set to the library’s maximum fragment length, ensuring true pairs are retained. The observed false-negative rate (real pairs filtered out) is below 1%.

2. Lightweight Alignment Algorithm

Filtered candidate pairs are aligned using a fast, bitwise approach that substitutes for conventional dynamic programming (DP). GenPairX observes that approximately 70% of read pairs deviate from the reference by only simple edits (mismatches or short indels).

Scoring parameters follow Minimap2’s affine-gap penalties:

• match: $(R_1, R_2)$8
• mismatch: $(R_1, R_2)$9
• gap open: $k$0
• gap extension: $k$1

Traditional DP (Needleman–Wunsch, Smith–Waterman) requires filling matrices $k$2, $k$3, $k$4:

$k$5

This incurs $k$6 time and space. GenPairX’s LightAlign instead computes the Hamming mask $k$7 (bitwise XOR; two bits per base) across possible indel shifts $k$8, then detects longest runs of 1's at sequence boundaries. This extraction of edit type, location, and score occurs in $k$9 time.

Smith–Waterman/Needleman–Wunsch requires $S_1 = \{s_1, s_2, s_3\}$0 time and space; GenPairX LightAlign operates in $S_1 = \{s_1, s_2, s_3\}$1 time (with $S_1 = \{s_1, s_2, s_3\}$2), $S_1 = \{s_1, s_2, s_3\}$3 space, and empirically solves $S_1 = \{s_1, s_2, s_3\}$4 of read pairs in $S_1 = \{s_1, s_2, s_3\}$5 cycles/read ($S_1 = \{s_1, s_2, s_3\}$6), compared to DP fallback at $S_1 = \{s_1, s_2, s_3\}$7 cycles/read.

3. Accelerator Architecture

GenPairX is implemented as a specialized ASIC with four pipelined modules:

Stage	Key Features	Throughput
Partitioned Seeding Module	6 parallel xxHash units, 2 GHz clock	333 M read-pairs/s/module
Near-Memory Seed Locator	32 HBM2 channels, sliding-window dispatch	192 M read-pairs/s at 1 GHz
Paired-Adjacency Filter	Dual-port SRAM FIFOs, single-cycle comparator	3 units to match NMSL
Light Alignment Module	Wide XOR datapath, parallel run finders	1.1 M pairs/s/unit, 174 units to match upstream

All modules reside on a 7 nm single ASIC die with bonded HBM2 stacks. Inter-module communication uses AXI-Stream links, and intermediate buffers manage burstiness and in-flight state.

4. Comparative Performance Analysis

GenPairX+GenDP (GenPairX front-end plus GenDP fallback) was benchmarked against Minimap2 on a Xeon CPU, BWA-MEM GPU, GenCache ASIC, and GenDP ASIC.

System	Throughput (Gbp/s)	Power (W)	Energy Efficiency (Gbp/s/W)	Area (mm²)	Area Efficiency (Gbp/s/mm²)
GenPairX+GenDP	277	209	1.32	381	0.73
GenDP	140	209	0.67	315.8	0.43
GenCache	2.17	11.2	0.19	33.7	0.06
BWA-MEM GPU (A100)	56	~300	0.19	815	0.07
Xeon CPU + Minimap2	0.037	~200	0.00019	300	0.00012

GenPairX+GenDP is approximately 1.43$S_1 = \{s_1, s_2, s_3\}$8 and 1575$S_1 = \{s_1, s_2, s_3\}$9 more energy efficient than GenCache and the CPU; 1.97$R_1$0 and 958$R_1$1 more area efficient, respectively.

End-to-end throughput figures:

• GenPairX+GenDP: 57.8 Gbp/s
• GenDP: 24.3 Gbp/s
• GenCache: 2.17 Gbp/s
• GPU: 0.056 Gbp/s
• CPU: 0.009 Gbp/s

5. Accuracy and Robustness

Variant calling benchmarks on 100$R_1$2 human whole-genome sequencing against the GIAB standard yield results for SNP and INDEL calling:

• Minimap2: SNP F$R_1$3 = 0.9913; INDEL F$R_1$4 = 0.9326
• GenPair+Minimap2 (no index filter): SNP F$R_1$5 = 0.9939/0.9887; INDEL F$R_1$6 = 0.9583/0.9300
• GenPair+Minimap2 (index filter threshold = 500): SNP F$R_1$7 = 0.9938/0.9887; INDEL F$R_1$8 = 0.9582/0.9299

The filtering heuristic with threshold = 500 yields a negligible impact on accuracy ($R_1$9F$S_2 = \{t_1, t_2, t_3\}$0 < 0.0001), with precision marginally higher and recall identical to Minimap2.

DP fallback rates:

• 2.09% of read-pairs require full DP (missed seeding)
• 8.79% require DP chaining/alignment (filtered out)
• 13.06% require DP alignment only

Thus, approximately 14% of pairs ever invoke heavyweight DP, bounding worst-case runtime and maintaining throughput stability.

GenPairX throughput remains stable at $S_2 = \{t_1, t_2, t_3\}$1192 M pairs/s for per-base error rates up to 0.2%. At 0.05% (Illumina HiFi), performance matches that for error-free data.

6. Technical Significance and Implications

GenPairX demonstrates that exploiting the paired-end insert-size window for joint seed-based filtering substantially increases the fraction of spurious mapping pairs eliminated prior to alignment, enhancing efficiency relative to single-read filtering. Lightweight, bitwise alignment obviates DP for the majority of read pairs. Specialized hardware modules and memory architecture maximize throughput, energy, and area efficiency while bounding worst-case computational cost through controlled DP fallback. The empirical preservation and slight enhancement of variant calling accuracy relative to widely used software mappers validates the practical reliability of this approach (Eudine et al., 27 Jan 2026).

A plausible implication is that future read-mapping pipelines can further benefit from architecture-aware codesign integrating joint filtering, efficient scoring, and modular accelerator pipelines.