When a sequencing lab cannot get enough reads on a few stubborn regions of the genome, it usually sequences everything more deeply. Those regions are typically exons with extreme guanine–cytosine (GC) content, where the short synthetic DNA "baits" that pull targets out of solution bind poorly. Wireless engineers faced a structurally similar problem and found a cheaper answer. A fixed antenna can sit in a "fading hole," a spot where reflected radio waves cancel. The old fix was to install many more fixed antennas. Movable-antenna systems instead shift a few elements a fraction of a wavelength out of the hole. This proposal carries that answer, as an optimization method rather than as hardware, into exome capture design. For each GC-refractory target, we choose where a small number of baits sit, including on moderate-GC flanking sequence within one DNA-fragment length of the target. The aim is to raise the worst-covered targets without adding baits or sequencing depth.
The opening exists because capture works at the level of DNA fragments, not probes. A bait hybridizes to one stretch of a roughly 200-base fragment and drags along the whole fragment, including any target sequence on it. A GC-rich exon that no bait can bind directly can therefore still be captured through a bait parked on a friendlier stretch nearby. That is the sequencing analogue of an antenna stepping out of a null. Which positions are friendly depends on the local GC landscape, on how far the bait sits from the target, and on how crowded the baits are. So the right positions come from an optimization problem, not a rule of thumb. The wireless literature has already worked out how to solve problems of this shape.
What the two literatures give us
The genomics side supplies the motivation and the measurement setting. A retrospective comparison of 746 tumors, Retrospective evaluation of whole exome and genome mutation calls in 746 cancer samples, found that whole-genome sequencing recovered 424 mutations in known cancer genes that exome sequencing missed. Missed loci include the high-GC gene VHL, and the authors attribute part of the discrepancy to loci of variable GC content. In PKD1, the GC-rich Exon 1 and the near-identical pseudogenes left 40 patients unsolved by short reads, as shown in Targeted long-read sequencing enables higher diagnostic yield of ADPKD by accurate PKD1 genetic analysis. That paper also reports that long reads raised yield in this sub-cohort from 62.5% to 87.5%. A cross-species capture study, Correction: An Effort to Use Human-Based Exome Capture Methods to Analyze Chimpanzee and Macaque Exomes, notes that extreme base composition depresses capture even with optimized probe sequences. These papers establish the gap. None of them treats bait position as a design variable.
The platform benchmark Performance comparison of four exome capture platforms on DNBSEQ-Series high throughput sequencer supplies the evaluation frame. It compares four commercial panels on the reference sample NA12878. It also documents the tension we target: panels trade capture specificity against coverage uniformity, and the shortfall is made up by over-sequencing. The simulator Comprehensive simulation of metagenomic sequencing data with non-uniform sampling distribution shows that sequencing efficiency follows a roughly quadratic curve in GC content, peaking near 40–50%. That curve is our starting response model, which we then replace with one fitted to data.
The antenna side supplies the method. Multiuser Communications with Movable-Antenna Base Station Via Antenna Position Optimization maximizes the minimum user rate over continuous antenna positions. It uses an outer particle-swarm search with a minimum-spacing penalty and an inner convex step, and reports a 63% gain in minimum rate over fixed antennas. Flexible-Position MIMO for Wireless Communications: Fundamentals, Challenges, and Future Directions shows that 4 antennas, each with 30 candidate positions, match 20 fixed antennas. That is the claim we hope to transfer: a few well-placed baits substitute for many uniformly tiled ones. Array Gain for Pinching-Antenna Systems (PASS) shows that packing elements too closely makes gain oscillate and drop. Our analogue is bait crowding: overlapping baits compete for the same fragments and increase off-target pulldown. We therefore keep an explicit minimum spacing.
We deliberately do not pursue the hardware readings the research file suggests. The file proposes steering terahertz fields with pinching antennas to unzip GC duplexes, building on DNA breathing dynamics in the presence of a terahertz field. That result is a one-dimensional model of a 64-base-pair chain. It says nothing about aqueous hybridization chambers. The file also proposes reconfigurable flow-cell sensors, extending End-to-End Optimization of High-Throughput DNA Sequencing. That paper's own model attributes yield loss to cluster interference, not to GC, so movable sensors would address the wrong bottleneck. Our position is that the antenna literature's transferable content is the optimization of where scarce, constrained elements sit under a spatially varying response. Bait placement is where that content matters in sequencing.
Core Claim
For a fixed total bait budget, jointly choosing each bait's position (including flanking offsets) and multiplicity under a learned capture-response model will substantially raise coverage of the worst-covered GC-extreme exons, compared with vendor tiling and with simply adding baits at default positions. Equivalently, the design will reach a given worst-target depth with fewer total reads. The claim is specific to repositioning. If extra baits at default positions do as well, the claim is false.
Let $P = \{p_1, \dots, p_B\}$ be bait start coordinates under a budget of $B$ baits. Let $c_t(P)$ be the expected depth on target $t$. Write $h(g_j, d_{jt})$ for the learned capture yield of a bait whose local GC fraction is $g_j$ and whose distance to target $t$ is $d_{jt}$; $h$ falls to zero beyond the fragment-length scale. Let $m(p_j)$ be the bait's mappability, which is zero inside pseudogene-identical sequence such as the PKD1P1–P6 regions. The design problem is
$$ \max_{P} \; \min_{t \in \mathcal{T}} \; c_t(P), \quad c_t(P) = \sum_{j=1}^{B} m(p_j)\, h(g_j, d_{jt}), \quad \text{s.t. } |p_i - p_j| \ge \delta \;\; \text{for } i \ne j, $$
where $\mathcal{T}$ is the set of refractory targets and $\delta$ is the minimum bait spacing. This mirrors the max-min, minimum-spacing structure of the movable-antenna problem. It also adds an off-target penalty that the antenna problem lacks.
Research Program
The first stage tests whether there is anything to exploit. The idea requires GC content to vary at sub-fragment scale around refractory exons, so that moderate-GC positions exist within reach. We will measure this directly on hg19 around the lowest-covered decile of exome targets. If the flanks are as GC-extreme as the exons, the idea fails at this stage.
The second stage learns the response function $h$. We regress per-base depth on bait GC, bait-to-base distance and local bait density, using public bait coordinate files for the four panels together with NA12878 exome data. A quadratic GC curve in the style of nuMetaSim is the null form. We test whether a flexible model predicts held-out targets better. Fitting across four panels with different tilings gives natural variation in bait placement, which is what identifies the distance term.
The third stage is the optimizer. Positions are effectively discrete at single-base resolution. Following Flexible-Position MIMO for Wireless Communications: Fundamentals, Challenges, and Future Directions, we restrict each bait to a candidate set within one fragment length of its target. We then solve with particle-swarm search plus a spacing penalty, as in Multiuser Communications with Movable-Antenna Base Station Via Antenna Position Optimization, and compare against an integer-programming relaxation to check that the heuristic is not leaving gains behind.
The fourth stage is a wet-lab spike-in panel on NA12878. It covers the refractory targets and adds the VHL and PKD1 Exon 1 loci, together with the exonic loci of the 424 WES-missed driver mutations in the 746-tumor comparison.
Baselines
The naive reference is the unmodified vendor exome, sequenced to the same total read count. We use Twist Exome 2.0 as the base, because the DNBSEQ benchmark reports it as the most uniform panel: 98.00% of bases at 30x and a FOLD_80_BASE_PENALTY of 1.31 (Performance comparison of four exome capture platforms on DNBSEQ-Series high throughput sequencer). FOLD_80_BASE_PENALTY is roughly the extra sequencing needed to bring 80% of bases up to mean depth. On worst-decile coverage specifically, we found no reported result on Lacuna for any panel. We will therefore measure it ourselves for every arm.
The strongest competitor is density boosting. It spends the same extra bait budget on the refractory targets but places the additional baits at standard tiling positions. This is roughly what vendors already do when they say probes are optimized. It is the baseline that isolates our key ingredient. To beat it, the position-optimized arm must give clearly higher worst-decile depth at matched bait count, matched reads and matched hybridization protocol, replicated across libraries.
The ablations are position-only (same bait count as vendor, repositioned) and multiplicity-only. We also replace the learned $h$ with the quadratic GC curve, to test whether the learned model is necessary. A computational baseline, LOESS GC normalization of the vendor data, is reported for variant-calling metrics.
Whole-genome sequencing is the ceiling on driver-locus recovery. The 746-tumor study is the reference for how much exome capture misses, and that study's whole-genome calls define the missed set.
Why not just…?
Why not just add more baits to the refractory exons? This is the real threat, and the density-boosting arm exists to settle it. If extra baits at default positions match repositioned ones, the extra yield is coming from multiplicity rather than placement, and we will report the idea as unnecessary. Our reason to expect a difference is mechanistic. Adding baits that sit on GC-extreme sequence multiplies a near-zero hybridization yield. A bait moved to a moderate-GC flank changes the yield itself. The PASS coupling result (Array Gain for Pinching-Antenna Systems (PASS)) also gives reason to expect diminishing or negative returns from crowding, a pattern we can check directly in the density arm.
Why not just correct coverage computationally? Normalization rescales the signal that exists. It cannot create reads where capture returned none, which is the regime behind the 424 missed driver mutations in Retrospective evaluation of whole exome and genome mutation calls in 746 cancer samples. The LOESS arm will show this by reporting variant sensitivity at those loci, which should not move under normalization.
Why not just use long reads or whole genomes? For PKD1, long reads are already the better tool, as Targeted long-read sequencing enables higher diagnostic yield of ADPKD by accurate PKD1 genetic analysis shows, and that paper itself notes the cost barrier. We do not claim to beat long reads on pseudogene-confounded loci. The mappability term in our objective simply avoids wasting baits there. Our target is high-throughput exome cohorts, where the question is whether cheap redesign closes part of the gap to whole-genome sequencing at exome cost.
Proposed Experiments
The in-silico experiment fits $h$ on three panels and predicts worst-decile coverage of the held-out fourth. It then compares optimized designs against vendor and density-boosted designs, measuring predicted reads needed to bring the worst decile of targets to 20x.
The decisive wet-lab experiment uses the spike-in panel on NA12878. It has the vendor, density-boost, position-only and full arms, each run in triplicate at matched reads on the same sequencer and the same one-hour hybridization protocol. Outcomes are worst-decile depth, FOLD_80_BASE_PENALTY, the fraction of bases at 30x, and capture specificity (off-target rate), since flanking baits may cost specificity. The success criterion is that the full design beats density boosting on worst-decile depth while losing at most a small, pre-registered amount of specificity.
A second wet-lab check reruns the same arms on the PancancerLight 800 reference standard. On that sample, the BOKE panel detected all 793 expected sites. The test is whether repositioning keeps that sensitivity while adding coverage at the GC-refractory driver loci.
The idea is falsified if any of three things happens: sub-fragment GC variation is absent in the first stage, the learned response does no better than the quadratic curve, or density boosting matches the full design.
References
- Retrospective evaluation of whole exome and genome mutation calls in 746 cancer samples
- Targeted long-read sequencing enables higher diagnostic yield of ADPKD by accurate PKD1 genetic analysis
- Correction: An Effort to Use Human-Based Exome Capture Methods to Analyze Chimpanzee and Macaque Exomes
- Performance comparison of four exome capture platforms on DNBSEQ-Series high throughput sequencer
- Comprehensive simulation of metagenomic sequencing data with non-uniform sampling distribution
- Multiuser Communications with Movable-Antenna Base Station Via Antenna Position Optimization
- Flexible-Position MIMO for Wireless Communications: Fundamentals, Challenges, and Future Directions
- Array Gain for Pinching-Antenna Systems (PASS)
- DNA breathing dynamics in the presence of a terahertz field
- End-to-End Optimization of High-Throughput DNA Sequencing