SpacerScope: binary-vectorized, genome-wide off-target profiling for RNA-guided nucleases without prior candidate-site bias Authors: Yanji Qu, Yaxuan Wang, Yan Wang, Haoru Tang, Qing Chen Published: 2026 ABSTRACT AbstractThe precision of CRISPR/Cas systems is fundamental to their application in plant and animal biotechnology. However, comprehensive sequence-based off-target candidate discovery remains a computational bottleneck, particularly in large and complex genomes. Here we developed SpacerScope, an off-target candidate discovery framework that enables unbiased, genome-wide discovery by leveraging binary vectorization, bitwise filtering, and right-end-anchored alignment. Benchmarking against human CIRCLE-seq data demonstrated that SpacerScope recovered 100% of validated off-target sites (6142/6142), matching the sensitivity of exhaustive algorithms. Crucially, SpacerScope achieved this maximum candidate recovery while substantially reducing computational overhead. In large-genome evaluations, SpacerScope maintained low peak memory usage of 2.20 GiB and achieved substantial runtime improvements over indel-aware comparator tools, including more than 50-fold speedup relative to Cas-OFFinder 3 (544 s versus 29 185 s). Furthermore, comparative analyses in polyploid species, such as the octoploid strawberry, revealed that SpacerScope identified larger sequence-compatible candidate burdens than standard web-based design platforms. Our results establish SpacerScope as a high-speed framework for sequence-based genome-wide off-target candidate discovery across diverse and highly repetitive genomic landscapes. The source code and program was publicly available at https://github.com/charlesqu666/SpacerScope. Short AbstractCRISPR/Cas sequence-based off-target candidate discovery remains computationally challenging in large, repetitive, and polyploid genomes. Existing tools either miss indel-containing candidate sites or incur prohibitive runtime and memory costs. We developed SpacerScope, a binary-vectorized framework that enables unbiased, genome-wide off-target candidate discovery without pre-selected candidate sites. By integrating bitwise filtering with right-end-anchored alignment, SpacerScope recovered 100% of validated off-target sites in human CIRCLE-seq data while using only 2.20 GiB of memory and achieving more than 10-fold speedup over indel-aware alternatives. Evaluation in plant genomes, including rice and octoploid strawberry, further demonstrated SpacerScope’s capacity to identify larger sequence-compatible candidate burdens overlooked by standard tools. SpacerScope thus provides a high-speed framework for sequence-based genome-wide off-target candidate discovery across diverse and highly repetitive genomic landscapes, supporting downstream prioritization. Introduction The diversification of Type II CRISPR/Cas systems has transformed genome engineering into a cornerstone of modern biotechnology. However, as the repertoire of CRISPR effectors and targeting rules expands, the demand for computational frameworks capable of high-precision target selection has grown increasingly acute. Although a variety of off-target search engines have been developed [1–3], balancing genome-wide sensitivity, indel-aware search, and computational scalability remains challenging in large and repetitive genomes. To cope with these challenges, existing algorithmic strategies were introduced, which generally fall into three categories, each with distinct trade-offs. Methods based on the Burrows-Wheeler Transform (BWT) or FM-index, such as Bowtie, offer efficient exact matching through compressed indexing but suffer from a precipitous decline in performance as sequence variation, particularly insertions and deletions (indels), increases [4]. Conversely, hashing and prefix-based indexing strategies, exemplified by FlashFry, facilitate rapid retrieval but can become memory-intensive on complex datasets [5]. Exhaustive or near-exhaustive tools such as the most recent Cas-OFFinder, CRISPRitz, and SWOffinder have expanded the detectable search space, including support for bulge- or indel-aware matching in some implementations, but this broader sensitivity often incurs substantial runtime costs in large genomes [1, 6, 7]. Moreover, the application of such tools to complex genomes presents additional challenges. In polyploid species characterized by high repeat content and extensive homeologous similarity, the number of plausible candidate sites can expand rapidly, exacerbating computational overhead and making complete candidate recovery difficult under practical runtime constraints. Indeed, recent evidence suggests that many single-nucleotide variants (SNVs) induced by base editors remain beyond the predictive reach of standard tools like Cas-OFFinder [8], posing significant safety concerns for crop engineering and therapeutic applications. A further complication arises from workflow-level candidate-site bias in current off-target assessment. In many practical studies, downstream validation is not initiated from an exhaustive genome-wide search, but from a restricted list of candidates pre-selected by web platforms such as CHOPCHOP [2] or CRISPOR [9]. Such pre-selection may underrepresent off-targets outside the assumed search model, including sites involving bulges or indels and highly similar target-gene homologues that are frequently encountered in plant genomes [10]. More generally, loci affected by PAM assumptions, repetitive or poorly mappable genomic contexts, or paralogous/homeologous sequence similarity may remain difficult to capture when candidate discovery is restricted to predefined prediction settings. As a result, the absence of a reported site may reflect an incomplete search space rather than a true lack of off-target potential. This distinction is biologically important because experimental validation, candidate prioritization, and safety interpretation are then all conditioned on an already truncated set of sites. Furthermore, the limited consensus in sgRNA selection and scoring across guide-design tools suggests that current predictive models are not yet robust enough for standalone use [11]. Biologically, this matters because a site omitted during candidate discovery could still become consequential if it lies in a dosage-sensitive gene, a regulatory element, or a highly similar duplicated locus that is subsequently carried forward into therapeutic or breeding decisions. To overcome these bottlenecks, we present SpacerScope, a high-performance computational framework designed to enhance the speed and fidelity of off-target analysis in complex eukaryotic genomes. SpacerScope integrates binary vectorization with a bitwise-operation-based pre-filtering mechanism to recover candidate sites genome-wide without requiring prior reduction to a model-derived shortlist. By employing 2-bit encoding and a right-end-anchored edit distance algorithm, SpacerScope enables exhaustive candidate discovery within user-defined search settings while reducing computational latency. Our benchmarking demonstrates that SpacerScope preserves candidate recovery under the tested conditions, avoiding the loss of potential target sites that can occur with heuristic candidate filtering and thereby providing a more complete basis for downstream validation and risk interpretation. Notably, SpacerScope is designed to resolve complex edit patterns in which substitutions and indels coexist while maintaining high candidate recovery and practical runtime performance in genome-scale searches. SpacerScope is written in C++ and is cross-platform compatible (Windows, Linux, and macOS). It also features a web-based interface for streamlined visualization, providing a versatile and robust solution for precision genome editing across diverse biological systems. Material and methods Binary-channel filtering Rather than relying on traditional four-character string representations (A, T, C, G), the encoding architecture of SpacerScope deconstructs each nucleotide sequence into four discrete binary channels, one for each nucleotide. For a sequence of length L, a position containing the target nucleotide is represented as 1 in its respective channel and 0 in the remaining three. For example, the sequence “ATCGATCG” is encoded as an A-channel bitset of “10001000,” with other channels populated analogously. By transforming character-based matching into integer comparisons and bitwise operations, the algorithm leverages low-level hardware optimizations to accelerate search speed. To accommodate insertions and deletions (indels) during scanning, the framework constructs an indel-aware variant filter for each query sequence within a user-defined edit distance. The algorithm systematically expands the query sequence to include all permissible substitutions, insertions, and deletions, generating a comprehensive set of permissible variants. To ensure uniform comparison across the reference genome, these variants, regardless of their actual length, are mapped into a fixed-length representation space. Shorter variants are normalized to maintain consistency with the canonical candidate-site length. To formalize the sensitivity guarantee of the binary-channel prefilter, let Q denote a spacer sequence and let k be the user-defined maximum edit-distance threshold. For each nucleotide channel b ∈ {A, C, G, T}, SpacerScope constructs a filter set Fb(Q, k) by enumerating all sequences that can be reached from Q through no more than k allowed edit operations, including substitutions, insertions, and deletions, under the same PAM orientation and fixed-length normalization scheme used during genome scanning. Let φb(S) denote the binary-channel encoding of a candidate sequence S in channel b after this normalization. A candidate target T is rejected by the prefilter only if φb(T) ∉ Fb(Q, k) for at least one nucleotide channel. If T is a true candidate within the specified threshold, namely if its right-end-anchored edit distance to Q satisfies danchor(Q, T) ≤ k, then there exists an edit path containing no more than k allowed operations that transforms Q into the normalized representation of T. This edit path is included during filter construction; therefore, φb(T) ∈ Fb(Q, k) for every nucleotide channel b. Consequently, no candidate satisfying danchor(Q, T) ≤ k can be discarded by the binary-channel prefilter. The filtering stage is therefore lossless with respect to sensitivity within the specified edit-distance threshold. Because the four channel memberships are tested independently, the prefilter may still retain candidates that do not correspond to a globally valid edit path. Such false-positive candidates are resolved by the subsequent right-end-anchored banded dynamic-programming validation. Specially, SpacerScope is specifically optimized for standard genomic bases (A, C, G, T), and ambiguous bases are handled explicitly according to the corresponding search path. We distinguish the degenerate symbol “N” used in a PAM pattern from ambiguous “N” bases present in input sequences. During PAM matching, “N” in the PAM pattern is treated as a degenerate base. For ambiguous bases in query or reference sequences, SpacerScope does not apply a general wildcard interpretation. In the substitution-only 2-bit scanning path, non-A/C/G/T characters are encoded using the default 00 state, which is the same 2-bit code as A. Therefore, in this fast scanning path, comparisons involving N are evaluated on the encoded representation: N versus A and N versus N are treated as equal, whereas N versus C, G, or T is treated as different. No additional ambiguity-aware rule is applied during substitution-only mismatch counting. In the indel-aware binary-channel filtering path, non-A/C/G/T bases are not assigned to any nucleotide-specific channel. Instead, the corresponding position has zero bits in all four A, C, G, and T channels. Thus, an ambiguous base does not contribute positive nucleotide identity information to the binary-channel filter, and a candidate containing such a base can pass the prefilter only if this all-zero channel state is compatible with the precomputed reachable channel patterns. During the final right-end-anchored DP validation and event-level post-processing, ambiguous bases are compared literally rather than treated as wildcards: N is considered a match only to N and a mismatch to A, C, G, or T. Ambiguous bases are also preserved as N during reverse-complement conversion and in the reported sequence context. Because ambiguous bases may reduce the interpretability of off-target calls, standard quality-control preprocessing of genome and query inputs is recommended before analysis. Right-end-anchored edit distance CRISPR/Cas-mediated cleavage is directionally constrained by the PAM. Because targeting specificity is typically highest near the PAM-proximal end, SpacerScope employs a right-end-anchored edit distance to quantify the divergence between a query sequence (Q) and a candidate target (T). While we use the term “right-end-anchored,” the algorithm dynamically adjusts orientation based on the PAM’s position relative to the spacer (5′ or 3′), which are defined by the users. This metric functions as a semi-global extension of the classical Levenshtein distance. Unlike standard global alignment, our approach permits a free offset at the PAM-distal (left) end. Prefixes that do not contribute to the optimal alignment are excluded from the edit cost, while the PAM-proximal (right) end remains strictly anchored. This asymmetric constraint effectively models length variations induced by indels while maintaining alignment fidelity in the biologically critical region adjacent to the PAM. By allowing sequences to initiate alignment from different starting points at the distal end, SpacerScope avoids inflating the edit distance with penalties caused by simple sequence offsets, focusing instead on true substitution and indel events. Computational workflow Low-complexity sequence filtering To mitigate non-specific binding and enhance computational throughput, candidate spacers undergo a user-definable low-complexity filtration step before genome-wide off-target searching. We implement a configurable DUST-like heuristic based on trinucleotide frequencies to identify and exclude low-complexity sgRNA candidates. For each trinucleotide t, a penalty score is calculated based on its occurrence nt. By default, this filter is enabled with a dust-threshold value of 12, which was shown to be able to balance the search speed and accuracy in our optimization test. Only sgRNA candidates with scores below the user-defined low-complexity threshold are retained, ensuring that the analyzed sgRNAs possess sufficient information density for unique genomic targeting. Importantly, this filtering step is applied only to sgRNA candidates before off-target searching. For sgRNAs that pass the filter, no genomic off-target candidate sites are excluded on the basis of DUST score during the subsequent search and validation stages. For analyses requiring off-target searches for all supplied sgRNAs, including low-complexity sgRNAs, this filtering process can be disabled by setting a sufficiently large dust-penalty value, such as 200. High-throughput substitution search For search scenarios restricted to base substitutions, SpacerScope utilizes a bitwise-parallel workflow leveraging compact 2-bit encoding. Both query and target sequences are mapped to binary representations (A = 00, C = 01, G = 10, T = 11), enabling a 20–30 bp spacer to be encapsulated within a single 64-bit unsigned integer. For targets on the antisense strand, the algorithm computes the reverse complement prior to encoding to maintain a unified search orientation. Mismatch quantification is performed via an optimized bitwise logic: An XOR operation identifies bit-level discrepancies between the query and target. Discrepancies are masked and compressed into a single-bit-per-nucleotide flag format. The total mismatch count is retrieved using the hardware-accelerated population count (popcount) instruction. By bypassing character-wise traversal in favor of basic bitwise primitives (XOR, SHIFT, MASK, and POPCOUNT), SpacerScope achieves near-theoretical maximum throughput for substitution scanning. To further maximize execution speed, we implemented a batch-scanning architecture. Rather than iterating through the reference genome for each individual query, which would incur prohibitive I/O overhead, SpacerScope groups multiple queries into a single processing batch. This allows the target set to be scanned sequentially, significantly improving cache locality and reducing redundant memory access. Pre-encoded targets are stored in contiguous arrays to ensure streamlined sequential reads, while the core comparison loops are unrolled to minimize branch-prediction penalties. The integrated substitution-search workflow is summarized in Fig. 1. Substitution-only search workflow in SpacerScope. (A) Data representation used for substitution search. Spacer and genomic target sequences are encoded into compact 2-bit nucleotide representations, with reverse-complement conversion applied to targets on the opposite strand. (B) Bitwise mismatch calculation for a single query-target pair. XOR, masking and population-count operations are used to convert nucleotide-level differences into a mismatch count without character-wise alignment. (C) Batch scanning strategy. Multiple query spacers are scanned against pre-encoded target collections in batches to improve cache locality, reduce redundant genome traversal and accelerate genome-wide substitution search. The workflow returns candidate off-target sites ranked by mismatch count and genomic location. Indel-aware search strategy Unlike substitution-only scanning, the inclusion of indels precludes the use of simple bitwise XOR for mismatch quantification. To maintain high-throughput performance without compromising sensitivity, SpacerScope employs a two-stage heuristic-to-exact pipeline: (i) high-speed binary-channel prefiltering to prune the search space, followed by (ii) Right-End-Anchored Banded Dynamic Programming (DP) for rigorous validation. The integrated indel-search workflow is illustrated in Fig. 2. Indel-aware search workflow in SpacerScope. The indel-search module combines binary-channel prefiltering with right-end-anchored banded dynamic programming. (A) Query and target sequences are decomposed into four nucleotide-specific binary channels and stored in a structure-of-arrays layout. (B) For each query, SpacerScope constructs filter sets representing sequence variants reachable within the user-defined edit-distance threshold. (C) Candidate genomic windows are retained only when all nucleotide channels satisfy the corresponding filter conditions. (D) Retained candidates are validated using banded dynamic programming, reducing the alignment space from the full matrix to a narrow edit-distance band. (E) The right-end-anchored constraint preserves the PAM-proximal end during alignment while allowing length variation at the PAM-distal end. (F) Early rejection and pruning rules terminate candidates whose length difference or partial dynamic-programming states exceed the allowed edit distance. The final output reports substitutions, insertions and deletions at event-level resolution. The implementation-level optimization panel summarizes the low-level engineering strategies used to improve throughput, including structure-of-arrays target storage, batch mask fusion for simultaneous query evaluation, and sequential scanning of the query table. The prefiltering stage abstracts genomic sequences into a one-hot bitset representation. Each sequence is decomposed into four discrete binary channels (A, C, G, and T) the same way as described above. This transformation converts the traditional string-matching problem into a multi-channel pattern recognition task. For each query, the algorithm generates all permissible variants, accounting for substitutions and indels, and maps them into a fixed-length channel space. This results in four comprehensive filter sets: FA, FC, FG, and FT. A candidate genomic site is retained for exact alignment only if its encoded values for all four channels belong to their respective filter sets: \documentclass[12pt]{minimal} \usepackage{amsmath} \usepackage{wasysym} \usepackage{amsfonts} \usepackage{amssymb} \usepackage{amsbsy} \usepackage{upgreek} \usepackage{mathrsfs} \setlength{\oddsidemargin}{-69pt} \begin{document} ${C}_i\in{F}_i,\mathrm{for}\ i\in \left\{A,C,G,T\right\}$\end{document}. Any candidate failing this membership test is immediately pruned. This design ensures that the computationally expensive DP alignment is reserved only for high-probability candidates, effectively pre-encoding the complexity of indels into a rapidly evaluable bitmask. Right-end-anchored banded dynamic programming Candidates surviving the prefiltering stage undergo exact validation via a specialized semi-global DP algorithm. To minimize the computational footprint, we implement a banded DP approach. Given a query of length m, a target of length n, and a maximum edit distance k, any candidate where |m - n| > k is discarded immediately. For eligible candidates, the algorithm computes only a diagonal band of width ~2 k + 1, reducing the time complexity from the standard O(mn) to O(mk). This validation stage is uniquely constrained by right-end anchoring, prioritizing alignment stability at the PAM-proximal end. We further optimize the DP execution through two early-termination rules: Immediate rejection if the cumulative length divergence exceeds the threshold k. Termination of the current alignment if all valid states within a given DP row exceed the maximum permissible edit distance. Consequently, the banded DP is executed over a highly restricted local state space, ensuring that “true” hits are validated with maximum precision and minimal latency. A detailed description of the dynamic-programming initialization, boundary conditions, and PAM-distal free-offset treatment is provided in Supplementary Note 1. Implementation of low-level optimizations To bridge the gap between algorithmic design and hardware performance, SpacerScope introduces several low-level optimizations including Structure of Arrays (SoA) Layout and Batch Mask Fusion. Genomic targets are stored in a contiguous SoA format by nucleotide channel. This layout maximizes spatial locality and minimizes cache misses during sequential genome-wide scans. When processing multiple queries simultaneously, individual channel filters are merged into unified lookup tables. These tables map channel values to query bitmasks, allowing the system to evaluate a batch of queries through a single set of bitwise intersections. This hardware-aware architecture, combined with our pruning heuristics, enables SpacerScope to perform exhaustive off-target discovery at speeds comparable to heuristic-based tools. In case of intense memory size requirements for vary large genome reference, we implemented tow modes, namely fasta and fasta-cut modes. These two modes use the same candidate-generation rules, binary-channel filtering strategy, right-end-anchored dynamic-programming validation, and post-processing procedures. Their difference lies only in how the reference genome is loaded and scanned. In fasta mode, the complete reference genome is loaded as a whole into memory and pre-encoded before genome-wide scanning. In fasta-cut mode, the reference genome is processed record by record, typically chromosome by chromosome or contig by contig. Each reference record is loaded, encoded, scanned, and released before the next record is processed, and the resulting candidate lists are then merged into the final output. Because candidate sites are searched within individual reference records and are not allowed to span different chromosomes or contigs, this record-wise processing does not change the searchable sequence space, but substantially reduces peak memory usage by avoiding simultaneous in-memory storage of the full reference genome. SpacerScope also supports annotation-based input modes, such as anno and anno-cut, in which gene sequences are first extracted from the genome according to user-provided gene identifiers before the same downstream search and validation procedures are applied. High-fidelity alignment and empirical risk scoring While the initial scanning stages identify potential query-target associations, these raw outputs represent low-resolution hits. To address this, SpacerScope implements a dedicated post-processing module to perform high-fidelity sequence parsing, heuristic-driven scoring, and event-level interpretation. First, to facilitate contextual sequence retrieval, the module consolidates raw candidates to calculate genomic occurrence frequencies. Because fixed-length search windows can artificially truncate alignments in the presence of indels, the engine retrieves the full genomic context, including flanking regions, directly from the reference assembly. Following sequence retrieval, SpacerScope utilizes a constrained event-enumeration procedure to evaluate the target site. The algorithm recursively explores all permissible interpretation paths, encompassing substitutions, insertions, and deletions within the defined edit-distance threshold. This exhaustive traversal ensures that the resulting alignment represents a biologically plausible interaction between the sgRNA and its genomic target, rather than a mere mathematical optimum. To assess the biological relevance of these alignments, candidate hits are prioritized using a heuristic empirical risk-scoring function derived from positional off-target trends summarized in the literature [12]. Specifically, we used the single-mismatch off-target frequencies reported for different PAM-proximal position intervals in Figure 4 of Reference [12], including positions 5–8, 9–12, and 13–16, to fit a smooth positional weighting curve for substitution events. This positional weighting was then empirically generalized to accommodate multiple mismatch events and extended to incorporate event type and event adjacency. For each non-perfect genomic match i, SpacerScope assigns a heuristic off-target risk score Pi, where 0 ≤ Pi ≤ 1. This value is intended to approximate the relative risk contribution of that candidate site within the SpacerScope ranking framework, but it should not be interpreted as an experimentally calibrated biological probability of editing. For a given sgRNA with n non-perfect genomic matches, the unique editing rate (UER) is calculated from these candidate-level risk scores as: UER represents a heuristic specificity index that summarizes the cumulative candidate-level risk scores across non-perfect genomic matches. Lower UER values indicate a larger cumulative sequence-compatible candidate burden within the SpacerScope scoring framework, whereas higher UER values indicate greater relative target uniqueness. Because the score is derived from literature-summarized positional trends and empirical extensions rather than from a dedicated statistical training model, UER should be used for relative comparison among candidate sgRNAs rather than as an experimentally calibrated absolute probability. Supported nuclease settings and search parameters SpacerScope is configurable at the sequence-search level for RNA-guided nuclease systems whose targeting rules can be represented by a contiguous guide/protospacer sequence together with an adjacent PAM. In the current command-line implementation, users can define the spacer length, primary PAM, optional alternate PAM, PAM orientation, mismatch threshold, and indel threshold. SpacerScope supports both 3′ PAM systems, such as SpCas9-like nucleases, and 5′ PAM systems, such as Cas12a-like nucleases, by allowing users to specify whether the PAM is located downstream or upstream of the spacer. In the current implementation, this setting controls spacer extraction and the normalization of guide-target orientation for downstream comparison and reporting. PAM degeneracy is supported through user-defined PAM patterns. In the current implementation, the symbol “N” in a PAM pattern is treated as a wildcard that can match any standard nucleotide. Other ambiguous IUPAC symbols are not interpreted as general PAM wildcards unless they are explicitly converted by the user before analysis. Only a single optional alternate PAM is supported directly by the current command-line interface. SpacerScope is designed for the short guide lengths typical of common RNA-guided nucleases. At the user-interface level, spacer length is user-defined, but the substitution-only search path uses a 2-bit encoding packed into a 64-bit integer and therefore directly supports guides up to 32 nt in that path. The indel-aware search path uses 64-bit binary-channel representations followed by right-end-anchored dynamic-programming validation and is likewise intended for short guide lengths rather than arbitrarily long protospacers. The benchmark analyses in this study mainly used SpCas9-like 20-nt guides with NGG/NAG PAM settings. For sequence differences, SpacerScope supports substitution-only searches as well as indel-aware searches containing substitutions, insertions, and deletions. Bulge-like guide-target differences are represented at the sequence-alignment level as insertion or deletion events rather than as a separate alignment class. Larger thresholds can be specified, but the indel-aware search space expands rapidly as the indel threshold increases, so practical analyses generally use small indel thresholds. Final report For classification and reporting, the post-processed hits are categorized into three distinct classes: perfect matches, substitution-only matches, or matches containing at least one indel. The final distance metric is updated to reflect the true number of divergence events identified during the refined alignment. These results are subsequently exported as a structured TSV table, which serves as the data source for an automated HTML visualization report. Unless otherwise stated, reported candidate counts refer to post-processed query-target-PAM records. Each record is defined by the sgRNA/query, reference sequence, genomic coordinate, strand, spacer sequence, PAM sequence, and refined alignment pattern. Exact duplicate records generated during intermediate scanning or result merging were collapsed during post-processing. Distinct genomic loci, strands, PAM instances, or sgRNA-query associations were retained as separate records, even when their genomic windows overlapped. Multi-copy hits at different genomic coordinates were also counted separately. Benchmark metric definitions For the CIRCLE-seq benchmark, candidate recovery was evaluated by coordinate overlap between tool-reported sites and experimentally detected CIRCLE-seq off-target sites for the corresponding sgRNA. A CIRCLE-seq site was considered recovered if at least one reported candidate overlapped the experimental site under the same coordinate-matching criterion used for the recall analysis. Recall was calculated as the number of recovered CIRCLE-seq sites divided by the total number of CIRCLE-seq validated sites. Apparent precision was calculated as the number of reported candidates overlapping CIRCLE-seq detected sites divided by the total number of reported candidates. Apparent F1 score was calculated as the harmonic mean of recall and apparent precision. These precision and F1 values are referred to as apparent metrics because CIRCLE-seq does not define a complete set of true negative genomic sites. Comparator benchmark analysis To compare with known relative tools including Cas-OFFinder 2, Cas-OFFinder 3, CRISPRitz, SWOffinder, and FlashFry, they were independently tested using the same reference genome assembly, sgRNA/query sequences, and PAM settings. All direct tool comparisons were performed on the same benchmark server running Ubuntu 22.04.5 LTS and equipped with dual AMD EPYC 7R32 processors, 96 physical CPU cores, and 512 GiB of physical RAM. Mismatch and indel/bulge thresholds were matched across tools whenever supported. When exact equivalence was not possible because of tool-specific parameter systems, the closest supported setting was used, and the corresponding indel-detection status was annotated in the relevant figure or table. All tools were run in single-threaded mode for direct comparison. Runtime was measured as wall-clock time, and peak memory was measured consistently using a process-tree profiling script that sampled the target command and its child processes. Software versions, genome assemblies, and indel/bulge settings are summarized in Supplementary Table S1. Key matching parameters, filtering steps, and post-processing settings used in the benchmark and workflow-level analyses are summarized in Supplementary Table S2. The complete benchmark archive, including raw inputs, retained outputs, run commands, profiling records, analysis and normalization scripts, expected output files, and source or summary data for the benchmark tables and figures, has been deposited in Zenodo under DOI 10.5281/zenodo.21188872. To avoid conflating different benchmark outcomes, we organized the Results according to the type of evidence being evaluated. Simulated datasets were used to assess algorithmic candidate recovery under known implanted ground truth. Rice WGS, human CIRCLE-seq, and GUIDE-seq datasets were used to assess recovery of experimentally detected off-target sites within defined search parameters. Runtime and memory usage were evaluated as computational performance metrics. Results Benchmarking via simulated genomic datasets To rigorously evaluate the algorithmic fidelity and sensitivity of SpacerScope across a spectrum of editing topologies, we performed validation using two simulated 1-Mb genomic datasets. These in silico benchmarks were engineered to include a ground-truth repertoire of heterogeneous mismatch profiles, encompassing single insertions coupled with substitutions, tandem insertions, multiple substitutions, and simultaneous insertion–deletion events. In Dataset A (low-frequency implantation), 413 ground-truth sites were embedded (~1 site per 2.5 kb) to simulate a sparse off-target landscape. In Dataset B (high-density implantation), 20 019 sites were embedded (~1 site per 50 bp) to assess algorithmic performance under extreme search-space saturation. Following analysis, SpacerScope successfully recovered 100% of the implanted ground-truth sites in both datasets (426 and 20 254 total hits detected, respectively), confirming the robustness of the bitwise-filtering and banded-DP pipeline across varied genomic densities (Fig. 3). Recovery of simulated off-target events across synthetic datasets. Different software was evaluated using two 1-Mb simulated genomic datasets containing implanted ground-truth off-target sites with different edit-event structures. (A) Dataset A represents a sparse off-target landscape. (B) Dataset B represents a high-density off-target landscape. Panels compare recovery rates for event classes including one insertion plus one substitution (1 ins + 1 sub), two insertions (2 ins), four substitutions (4 subs) and one insertion plus one deletion (1 ins + 1 del). Bars indicate the percentage of implanted sites recovered by each tool; numbers above bars indicate recovered sites relative to the total number of implanted sites in that category. Tool labels indicate whether indel detection was enabled or unsupported under the benchmark settings. Ins, insertion; Del, deletion; sub, substitution. Events are defined relative to the target. All tested software was forced to run in single-threaded mode to ensure fairness. For the comparative analysis, we benchmarked the same synthetic datasets against established mainstream tools. The results revealed clear sensitivity deficits in Cas-OFFinder 2, Cas-OFFinder 3, and FlashFry. Notably, although Cas-OFFinder 3 claims to support non-canonical indel alignments, it still failed to recover a substantial proportion of the ground-truth sites (Fig. 3). These findings highlight the superior analytical resolution of SpacerScope, especially its ability to resolve complex edit patterns that remain difficult for current software to detect. Validation via empirically verified off-target sites in Oryza sativa To evaluate recovery of previously reported off-target sites in a complex crop genome, we leveraged a high-confidence off-target dataset from a previous whole-genome sequencing (WGS) study in rice (O. sativa) [13]. Using a previously targeted sequence as a query (Os02circ25329 locus: 5′-GCAGCTCTGACATGTGGGCCCGG-3′), we screened the rice reference genome with SpacerScope, applying a threshold of 4 mismatches and 2 indels. SpacerScope achieved 100% sensitivity by successfully recovering all 12 previously validated off-target sites. The UER, a heuristic specificity index integrated into SpacerScope, was calculated at 0.56% for this sgRNA, consistent with the extensive off-target activity observed experimentally. We further extended our analysis to the remaining 11 sgRNAs described in the same study. For these guides, the authors reported negligible off-targeting or variations attributable to background genetic noise rather than Cas9 activity. Consistent with these biological observations, SpacerScope assigned 7 of these sgRNAs a UER exceeding 90%, with the lowest recorded value at 64.76%. These results suggest that SpacerScope’s heuristic specificity metric is broadly consistent with the experimentally observed off-target patterns in this rice dataset. A comprehensive summary of predictions for rice targets is provided in Table 1. Table 1Off-target prediction summary for rice sgRNAs.TotalExactMismatchIndelUnique editing rate (UER)Os03circ00204-sgRNA0214112196.41%Os03circ00204-sgRNA0119117191.53%Os02circ25329-sgRNA0220118192.98%Cas9-B:29127193.80%Cas9-F32227391.97%Cas9-G40236296.52%Cas9-D53245673.89%Cas9-E57250571.09%Cas9-H63161190.91%Cas9-C17911651364.76%Os02circ25329-sgRNA011805152230.56%Cas9-A1871184278.46%Note: Total indicates the number of candidate genomic sites reported by SpacerScope. Exact, Mismatch and Indel indicate candidates classified as perfect matches, substitution-only matches and matches containing at least one insertion or deletion, respectively. Unique Editing Rate (UER) is a heuristic specificity index derived from the candidate-level empirical risk scores of all non-perfect genomic matches, as defined in the High-Fidelity Alignment and Empirical Risk Scoring section. UER should not be interpreted as an experimentally calibrated editing probability. Off-target prediction summary for rice sgRNAs. Note: Total indicates the number of candidate genomic sites reported by SpacerScope. Exact, Mismatch and Indel indicate candidates classified as perfect matches, substitution-only matches and matches containing at least one insertion or deletion, respectively. Unique Editing Rate (UER) is a heuristic specificity index derived from the candidate-level empirical risk scores of all non-perfect genomic matches, as defined in the High-Fidelity Alignment and Empirical Risk Scoring section. UER should not be interpreted as an experimentally calibrated editing probability. We next benchmarked SpacerScope against CRISPRitz, SWOffinder, and FlashFry on the rice genome. All tools were executed in single-threaded mode to ensure a fair comparison of algorithmic efficiency. Because the 12 previously validated off-target sites in rice [13] were characterized by low sequence divergence (substitution-only), all five tools successfully recovered these primary hits. However, across the 12 tested sgRNAs, SpacerScope consistently identified a broader range of potential off-target candidates in 11 cases (Table 2). This expanded detection capability is primarily attributed to SpacerScope’s superior capacity to resolve indel-inclusive alignments, which are frequently overlooked by conventional heuristic-based or substitution-only tools. Table 2Comparative benchmarking of sgRNA targeting in the rice genome.SpacerScope (fasta mode)SpacerScope (fasta-cut mode)Cas-OFFinder 2Cas-OFFinder 3FlashFryCRISPRitzSWOffinderTotal detected sites873873832869832832860Os03circ00204-sgRNA0214141314131314Os03circ00204-sgRNA0119191819181818Os02circ25329-sgRNA0220201920191920Cas9-B29292829282829Cas9-F32322932292931Cas9-G40403940393940Cas9-D53535252525253Cas9-E57575757575757Cas9-H63636263626262Cas9-C179179173177173173179Os02circ25329-sgRNA01180180157180157157157Cas9-A187187185186185185186Peak memory (GiB)3.450.371.211.212.121.080.91Runtime (s)656248166163245811Indel detectionEnabledEnabledNot supportedEnabledNot supportedAvailable but not enabledEnabledNote: Values for individual sgRNAs indicate the number of candidate sites detected by each tool. All tools were run in single-threaded mode under comparable search thresholds where supported. SpacerScope was evaluated in fasta and fasta-cut modes. All tested software was forced to run in single-threaded mode to ensure fairness. Comparative benchmarking of sgRNA targeting in the rice genome. Note: Values for individual sgRNAs indicate the number of candidate sites detected by each tool. All tools were run in single-threaded mode under comparable search thresholds where supported. SpacerScope was evaluated in fasta and fasta-cut modes. All tested software was forced to run in single-threaded mode to ensure fairness. While Cas-OFFinder 2 and CRISPRitz were marginally faster (48 s and 45 s, respectively), they were restricted to a “substitution-only” mode for this benchmark; their inherent inability to detect complex indels limits their utility in high-precision applications. It is noteworthy that while CRISPRitz technically supports indel detection, its computational efficiency in this mode is prohibitively low in our test, leading to extreme processing latencies when handling complex genomes. Consequently, it was evaluated here only in substitution-matching test. In contrast, among the subset of tools actively performing indel detection, SpacerScope (FASTA-cut mode) demonstrated a decisive efficiency advantage (62 s), significantly outperforming SWOffinder (811 s) and Cas-OFFinder 3 (1661 s). These results underscore the robustness and speed of our bitwise-parallel architecture and pre-filtering mechanisms when resolving complex editing topologies. Sensitivity benchmarking against CIRCLE-seq experimental data To evaluate the reliability of SpacerScope in a large-scale and complex genomic context, we utilized the human CIRCLE-seq dataset [14] as a representative in vitro benchmark for off-target assessment. SpacerScope was configured with inclusive parameters (six mismatches, two indels) to target both canonical (NGG) and non-canonical (NAG) PAM sites. As shown in Fig. 4, SpacerScope demonstrated a strong ability to balance sensitivity and performance. Within the 6142 CIRCLE-seq validated sites, both SpacerScope modes and SWOffinder achieved zero false negatives, recovering 100% of the true off-targets. FlashFry, CRISPRitz, Cas-OFFinder 3, and Cas-OFFinder 2 recovered 6139 sites, corresponding to a recall of 0.9995. The total candidate counts also differed substantially across tools, reflecting the practical review burden associated with each search strategy. SpacerScope reported 437 146 candidates in both fasta and fasta-cut modes. SWOffinder reported a comparable number of candidates (437,143), while CRISPRitz, Cas-OFFinder 3, Cas-OFFinder 2 and FlashFry reported 437 041, 437 048, 437 041, and 437 041 candidates, respectively. Because CRISPRitz exhibited substantial slowdown when indel detection was enabled in complex genomes, its indel-aware mode was not enabled in this comparison. Benchmarking against CIRCLE-seq data using SpacerScope. SpacerScope was evaluated in both fasta and fasta-cut modes with parameters allowing up to six mismatches and two indels, including NGG and NAG PAMs. Comparator tools were configured so that their target matching ranges were aligned with those of SpacerScope as closely as possible. Total candidates indicate the number of post-processed query-target-PAM records returned by each tool under the benchmark search settings. Recovered sites indicate the number of experimentally detected CIRCLE-seq off-target sites recovered by each tool. Recall was calculated using 6142 CIRCLE-seq validated sites as the denominator. Apparent precision and apparent F1 score were calculated using CIRCLE-seq detected sites as assay-defined positives. Because CIRCLE-seq does not provide a complete set of true negative genomic sites, these values should be interpreted as CIRCLE-seq-defined apparent metrics reflecting candidate-list compactness rather than absolute biological specificity. Peak memory was measured as peak proportional set size in GiB, and runtime was measured as wall-clock time in seconds. The indel-detection labels indicate whether indel-aware search was enabled, available but not enabled, or not supported under the benchmark configuration. To complement the recall-based comparison, we further calculated CIRCLE-seq-defined apparent precision and apparent F1 score for each tool. SpacerScope reported 437 146 total candidates in both fasta and fasta-cut modes, resulting in an apparent precision of 1.41% and an apparent F1 score of 0.0277. SWOffinder showed same apparent precision value, whereas CRISPRitz, Cas-OFFinder 3, Cas-OFFinder 2, and FlashFry showed apparent precision values of 1.40%; all tools showed an apparent F1 score of 0.0277 under this benchmark configuration. These results indicate that SpacerScope achieved comparable detection precision with those popular tools. In terms of resource efficiency, SpacerScope (FASTA-cut mode) and SWOffinder were both superior, with memory consumption of only 2.20 GiB, significantly lower than the Cas-OFFinder series (9.60 GiB). Furthermore, SpacerScope maintained a clear edge in processing speed: the FASTA-cut mode completed the task in 544 s, whereas SWOffinder, despite its high sensitivity and low memory footprint, required 6273 s. Cas-OFFinder 3 required 29 185 s under the same benchmark framework, demonstrating substantially higher computational cost. We also performed an additional GUIDE-seq recovery analysis using a subset from the CRISPR-Bulge dataset (Supplementary Table S3) [15]. This subset contained 205 experimentally detected off-target sites, of which 196 were within the predefined search space used in this benchmark. SpacerScope recovered all 196 in-parameter-space sites in both fasta and fasta-cut modes, corresponding to a parameter-space recall of 100.0%. The remaining nine experimentally detected sites were outside the predefined search space, including five PAM-mismatch sites and four sites exceeding the mismatch threshold. Interestingly, the bulge-annotated sites in this subset could also be explained by substitution-only matching within the 20 bp protospacer window under the selected mismatch threshold in SpacerScope. Together, these results show that SpacerScope, particularly in FASTA-cut mode, maintains complete recovery of CIRCLE-seq validated sites while substantially reducing memory usage and runtime compared with the main indel-aware comparator tools. This combination of validated-site recovery and computational efficiency supports the use of SpacerScope for sequence-based off-target candidate discovery, while biological editability still requires context-specific experimental validation. Workflow-level demonstrations using web-based design platforms To evaluate the practical utility of SpacerScope in established workflows, we conducted workflow-level comparisons with CHOPCHOP v2 [2] and CRISPOR [9]. These analyses were not intended as direct algorithmic benchmarks, because identical input regions, assemblies, PAM settings, mismatch thresholds, and filtering rules could not be fully matched across platforms. We selected four representative loci across mammalian and plant models: EXT1 (human and mouse), AGAMOUS (Arabidopsis), and its rice ortholog OsMADS3. These genes were selected for their fundamental roles in skeletal development (EXT1) and reproductive fitness (AGAMOUS/OsMADS3), representing high-priority targets for both therapeutic and agricultural engineering [16, 17]. Using the default CHOPCHOP v2 workflow, we identified the top-ranked sgRNA for each locus and re-evaluated their sequence-compatible off-target candidate profiles using SpacerScope. Our analysis revealed that SpacerScope consistently identified a greater number of potential off-target sites, particularly as sequence divergence increased (Table 3). For instance, in the human EXT1 locus, CHOPCHOP identified no off-targets within its maximum threshold (MM < 3), whereas SpacerScope detected four potential sites. This discrepancy illustrates that candidate lists generated under default web-platform settings can differ from those obtained by a broader SpacerScope genome-wide sequence search. Similar observations were obtained when re-analyzing CRISPOR-derived guide-design outputs [9]. Due to CRISPOR’s input length restrictions (<2300 bp), only partial gene fragments were used for comparative analysis. Despite the use of same reference assemblies, SpacerScope identified significantly more potential off-target loci, often an order of magnitude higher in the MM4 category (Table 4). These findings suggest that default web-based workflows may return narrower candidate lists under their own input constraints, genome assemblies, mismatch thresholds, and filtering rules. Table 3Representative gRNA case comparisons between SpacerScope and CHOPCHOP across different species.EXT1(Homo sapiens)EXT1(Mus musculus)AGAMOUSOsMADS3SpacerScope(MM0/MM1/MM2/MM3)0/1/2/1(CTTATATCACGTCCATAACGGGG)0/1/0/0(ATATCACGTCCATAACGGGGAGG)0/0/0/0(GGACAATTCTAACACCGGATCGG)0/1/3/0(GCAGCCACTGAACATGATGGGGG)CHOPCHOP(MM0/MM1/MM2/MM3)0/0/0/0(CTTATATCACGTCCATAACGGGG)0/0/0/0(ATATCACGTCCATAACGGGGAGG)0/0/0/0(GGACAATTCTAACACCGGATCGG)0/0/0/0(GCAGCCACTGAACATGATGGGGG)Note: Values are shown as MM0/MM1/MM2/MM3, indicating the number of candidate sites with 0, 1, 2, or 3 mismatches. The sequence in parentheses is the sgRNA target sequence used for comparison. Because CHOPCHOP supports a maximum mismatch threshold of 3, SpacerScope results are restricted to the same mismatch range in this table. Representative gRNA case comparisons between SpacerScope and CHOPCHOP across different species. Note: Values are shown as MM0/MM1/MM2/MM3, indicating the number of candidate sites with 0, 1, 2, or 3 mismatches. The sequence in parentheses is the sgRNA target sequence used for comparison. Because CHOPCHOP supports a maximum mismatch threshold of 3, SpacerScope results are restricted to the same mismatch range in this table. Table 4Representative gRNA case comparisons between SpacerScope and CRISPOR across different species.EXT1_part(Homo sapiens)EXT1_part(Mus musculus)AGAMOUS_partOsMADS3_partSpacerScope(MM0/MM1/MM2/MM3/MM4)0/1/2/15/285(TAGTAATAGTTACAACCGAAAGG)0/0/0/6/69(GATGATCGAGCCACGGCCCTCGG)0/1/1/0/12(GACCTGAAATTGATATAGACAGG)0/0/0/6/25(TAAATATCAACTTGCCTAACTGG)CRISPOR(MM0/MM1/MM2/MM3/MM4)0/0/0/6/103(TAGTAATAGTTACAACCGAAAGG)0/0/0/3/28(GATGATCGAGCCACGGCCCTCGG)0/0/0/0/3(GACCTGAAATTGATATAGACAGG)0/0/0/0/3(TAAATATCAACTTGCCTAACTGG)Note: Values are shown as MM0/MM1/MM2/MM3/MM4, indicating the number of candidate sites with 0 to 4 mismatches. The sequence in parentheses is the sgRNA target sequence and PAM used for comparison. Because CRISPOR accepts input sequences shorter than 2300 bp, only partial gene fragments were used. Representative gRNA case comparisons between SpacerScope and CRISPOR across different species. Note: Values are shown as MM0/MM1/MM2/MM3/MM4, indicating the number of candidate sites with 0 to 4 mismatches. The sequence in parentheses is the sgRNA target sequence and PAM used for comparison. Because CRISPOR accepts input sequences shorter than 2300 bp, only partial gene fragments were used. Workflow-level re-analysis in a polyploid genome Off-target candidate discovery is particularly challenging in polyploid species due to high homeologous sequence similarity. We performed a workflow-level re-analysis of sgRNAs selected by CRISPR-PLANT v2 [18] using the MSH1 gene (a regulator of organellar genomic stability) [19] as an example in the octoploid strawberry (Fragaria × ananassa cv. Benihoppe) genome. This analysis was not intended as a direct algorithmic benchmark, because CRISPR-PLANT v2 and SpacerScope differ in candidate identification, scoring systems and candidate-ranking criteria. Beyond the differences in time consumption, our analysis identified a notable discrepancy in how targets were categorized. Several candidate sgRNAs assigned the highest recommendation rating (“A1”) by CRISPR-PLANT v2 were found by SpacerScope to have a UER of less than 70% in the SpacerScope prioritization framework (Table 5). This suggests that re-analysis with a broader genome-wide sequence search can reveal additional sequence-compatible candidates in the complex strawberry genome. These results illustrate the value of using SpacerScope as a complementary re-analysis tool when navigating the redundant genomic landscapes of polyploid crops. Notably, because no independent strawberry off-target validation dataset was available for these MSH1 candidates, this analysis should be interpreted as a heuristic prioritization comparison rather than as experimental evidence of differential off-target activity. Table 5Overlap between SpacerScope MSH1 matching results and CRISPR-PLANT v2 ratings in Fragaria × ananassa.CountA0A1A0.1B0B1B0.1UER > 90% (SpacerScope results)11000220UER 70%–89% (SpacerScope results)41000111UER 50%–69% (SpacerScope results)58010476UER < 50% (SpacerScope results)1230007116Note: A0, A1, A0.1, B0, B1, and B0.1 denote the numbers of candidate sites assigned to the corresponding CRISPR-PLANT v2 rating categories. Unique editing rate (UER) groups are based on the heuristic specificity index calculated from the candidate-level empirical risk scores of all non-perfect genomic matches, as defined in the high-fidelity alignment and empirical risk scoring section. These groupings are intended only for heuristic comparison and downstream prioritization; they do not represent independent strawberry off-target validation or experimentally calibrated absolute editing probabilities. Overlap between SpacerScope MSH1 matching results and CRISPR-PLANT v2 ratings in Fragaria × ananassa. Note: A0, A1, A0.1, B0, B1, and B0.1 denote the numbers of candidate sites assigned to the corresponding CRISPR-PLANT v2 rating categories. Unique editing rate (UER) groups are based on the heuristic specificity index calculated from the candidate-level empirical risk scores of all non-perfect genomic matches, as defined in the high-fidelity alignment and empirical risk scoring section. These groupings are intended only for heuristic comparison and downstream prioritization; they do not represent independent strawberry off-target validation or experimentally calibrated absolute editing probabilities. Computational scaling and performance benchmarking To assess the resource footprint and scalability of SpacerScope, we performed multi-threaded benchmarks using full-length AGAMOUS and EMX1 sequences against the Arabidopsis and human genomes, respectively. Benchmarking was performed on the same server described in the Comparator Benchmark Analysis section. SpacerScope demonstrated robust computational scaling, with runtimes decreasing significantly as thread counts increased (Fig. 5A). In the human genome scan, SpacerScope maintained high throughput even when exhaustive indel-search was enabled, a testament to the efficiency of the bitwise pre-filtering architecture. Peak memory usage was observed to be stable and predictable across varied workloads (Fig. 5B). Among these results, the speed improvement at 64 threads was relatively modest compared to 32 threads, which is attributable to the limited memory bandwidth of the test machine. The fasta-cut mode performed substantially worse than the fasta mode under multi-threading, as the fasta-cut mode was originally designed to enable computation on machines with restricted RAM. Future software updates will optimize the fasta-cut mode to achieve exceptionally high speed while maintaining low memory footprint. These results establish SpacerScope as a high-efficiency solution capable of handling large-scale genomic data with a manageable resource footprint, making it suitable for both institutional servers and individual research workstations. Runtime and peak memory usage of SpacerScope across genome-scale workloads. (A) Runtime of SpacerScope using 4, 32, and 64 threads across four genome-scale workloads: analysis of the AGAMOUS gene in Arabidopsis using fasta mode, analysis of the AGAMOUS gene using fasta-cut mode, analysis of the EMX1 gene in the human genome using fasta mode, and analysis of the EMX1 gene using fasta-cut mode. The y-axis shows execution time in seconds on a log10 scale. (B) Corresponding system-wide peak memory usage for the same workloads and thread counts. Benchmarks were performed on a server running Ubuntu 22.04.5 LTS and equipped with dual AMD EPYC 7R32 processors, 96 physical CPU cores, and 512 GiB of physical RAM. Discussion The precision of CRISPR/Cas systems remains the cornerstone of their utility in functional genomics and crop improvement. Existing off-target analysis tools often face a trade-off between computational efficiency and detection sensitivity, and current validation workflows can also be biased by their dependence on predicted candidate sites [10, 11]. SpacerScope was designed to address both of these limitations by combining high-throughput binary vectorization with unbiased, genome-wide analysis that does not depend on a narrowly preselected candidate set. Importantly, the objective of this study is to improve the completeness and efficiency of sequence-based off-target candidate discovery, rather than to directly predict editing activity, cellular off-target frequency, or clinical safety. Algorithmic efficiency via state-space compression A key innovation of SpacerScope is the implementation of binary-channel pre-filtering. Conventional whole-genome off-target search tools such as Cas-OFFinder perform direct sequence comparison under predefined mismatch or bulge constraints [1], and approximate matching generally becomes more expensive as the allowed edit space and candidate volume increase [20]. SpacerScope instead adopts the general logic of bit-parallel approximate matching [21] but adapts it to CRISPR off-target search through binary-channel encoding of nucleotides. This representation reduces the effective state space before downstream alignment and enables aggressive pruning of low-probability loci without sacrificing sensitivity. Compared with tools that support indel detection, such as SWOffinder, SpacerScope achieves a substantial leap in detection speed. Engineering optima: memory bandwidth versus arithmetic complexity In developing our substitution-detection module, we evaluated two broad classes of approximate-matching strategies: q-gram-based filtering [22] and bit-parallel matching [21]. Our benchmarks showed that, for short spacers scanned against very large target libraries, the dominant bottleneck is memory bandwidth and cache locality rather than the arithmetic cost of Hamming-distance calculation itself. Under these conditions, batch brute-force matching with compact 2-bit encoding provided the best engineering trade-off: it minimizes per-target memory footprint, preserves sequential access patterns, and avoids the index-construction and reuse assumptions required by q-gram-based methods. We therefore retained 2-bit-encoded batch scanning as the main substitution-detection path in SpacerScope. Mitigating analysis bias and indel blind spots A significant contribution of this work is the reduction of candidate-discovery bias in sequence-based off-target analysis. Plant off-target studies have noted that validation is often restricted to predicted candidate sites rather than unbiased genome-wide analysis [9], which can create a blind spot for sequence-compatible off-target candidates in genomic regions not included in the predefined validation list. In our benchmarks, established tools such as Cas-OFFinder and web-based platforms such as CHOPCHOP showed limited recovery of complex mismatch patterns, including tandem insertions and combined substitution-indel events [1, 2]. SpacerScope addresses this gap with a right-end-anchored edit distance that is motivated by the PAM-proximal asymmetry of CRISPR targeting [23, 24]. By allowing flexible trimming at the PAM-distal side while preserving the PAM-proximal anchor, the method can accommodate length variation patterns that remain biologically relevant in bulge-containing off-targets [25]. We then refine the interpretation of each retained site with a heuristic empirical risk-scoring framework derived from literature-summarized positional off-target trends and conceptually related to data-driven off-target scoring models [26], yielding an interpretable candidate-level prioritization index rather than a raw distance value alone, without treating this index as a direct estimate of cellular editing probability. Validation and future perspectives The robustness of this framework is supported by our retrospective analysis of CIRCLE-seq data [14], in which SpacerScope recovered 100% of validated sites within the defined parameter space. The tool also identified additional sequence-compatible candidates that were initially misclassified because of more complex edit patterns, highlighting the corrective value of the refined-alignment stage. It should be noted, however, that CIRCLE-seq is an in vitro assay and does not fully capture the cellular determinants of CRISPR/Cas off-target activity. Factors such as chromatin accessibility, local DNA repair bias, cell-type-specific nuclease activity, DNA damage response, and epigenomic context can influence whether a sequence-compatible off-target site is actually edited in a cellular context [27]. Therefore, the CIRCLE-seq benchmark should be interpreted primarily as an evaluation of SpacerScope’s ability to recover sequence-defined off-target candidates within a controlled search space, rather than as a complete model of cellular off-target editing outcomes. A conventional biological false-positive rate was not calculated for this benchmark because it would require a well-defined set of true negative genomic sites. CIRCLE-seq provides experimentally detected positives under a specific in vitro condition, but the absence of a candidate from the CIRCLE-seq detected set does not prove that the site lacks off-target potential under other experimental or cellular contexts. Therefore, non-CIRCLE-seq-detected candidates were not interpreted as definitive biological false positives. Instead, apparent precision and apparent F1 score were reported as assay-defined indicators of candidate-list compactness and downstream validation burden. Computational trade-offs nevertheless remain. When the allowed indel threshold is increased substantially, variant generation grows non-linearly and runtime rises accordingly. In bulge-aware off-target search, practical analyses commonly restrict the search space to one or two bulges [5], and our own rice and CIRCLE-seq benchmarks indicate that an indel threshold of 2 was sufficient under the settings evaluated here. Future iterations could therefore focus on more aggressive pruning heuristics or GPU-accelerated variant generation to expand the searchable parameter space without sacrificing sensitivity. Incorporating cell-based or in vivo datasets, such as GUIDE-seq or DISCOVER-seq, together with chromatin-aware or repair-aware computational noise models, will also be an important direction for future development and validation. Although SpacerScope supports user-defined PAM orientation, PAM patterns, spacer lengths, mismatch thresholds, and indel thresholds at the sequence-search level, the benchmark analyses in this study were mainly performed under SpCas9-like 20-nt guide and NGG/NAG PAM settings. Therefore, additional nuclease-specific benchmarking will be needed before extending the performance conclusions to other RNA-guided nuclease classes. Key Points SpacerScope provides a binary-vectorized framework for genome-wide off-target profiling of RNA-guided nucleases without relying on pre-selected candidate sites. SpacerScope provides a binary-vectorized framework for genome-wide off-target profiling of RNA-guided nucleases without relying on pre-selected candidate sites. SpacerScope supports both substitution-only and indel-aware searches, combining fast bitwise scanning with right-end-anchored alignment to improve sensitivity in complex genomes. SpacerScope supports both substitution-only and indel-aware searches, combining fast bitwise scanning with right-end-anchored alignment to improve sensitivity in complex genomes. Benchmarking against simulated datasets, rice whole-genome sequencing data, and human CIRCLE-seq data showed that SpacerScope recovered validated off-target sites with high sensitivity. Benchmarking against simulated datasets, rice whole-genome sequencing data, and human CIRCLE-seq data showed that SpacerScope recovered validated off-target sites with high sensitivity. Comparative analyses indicate that commonly used web-based design platforms may underestimate off-target risks, especially when sequence divergence or indels are present. Comparative analyses indicate that commonly used web-based design platforms may underestimate off-target risks, especially when sequence divergence or indels are present. SpacerScope is cross-platform, provides a web-based interface, and is suitable for off-target assessment in large, repetitive, and polyploid genomes. SpacerScope is cross-platform, provides a web-based interface, and is suitable for off-target assessment in large, repetitive, and polyploid genomes. Supplementary Material