Phasis: a software tool for register-resolved discovery of plant phased small RNA loci

Abstract

Plant PHAS locus discovery remains challenging because phasiRNA-producing loci must be distinguished from other sRNA-producing regions with high abundance or apparent periodicity. This problem is especially acute for reproductive 24-PHAS loci, which occur within genomes that also produce abundant 24-nt siRNAs from nonPHAS regions. We present Phasis, an open-source Python software tool for plant PHAS-locus discovery from small RNA sequencing data. Phasis combines statistical evidence for phased accumulation with locus-level features and a Register-Resolved Locus Interpretation Layer that evaluates whether candidate loci show coherent phased architecture. Across diverse plant datasets, Phasis recovered validated or annotated 21- and 24-PHAS loci with a strong balance between call-level precision and reference-locus recall, and generally outperformed PhaseTank and ShortStack in matched benchmark analyses. The register-resolved interpretation layer reduced unsupported calls by separating coherent phased loci from ambiguous sRNA-producing regions. In maize dcl5 mutant libraries, Phasis showed strong depletion of 24-PHAS recovery, supporting DCL5-dependent recovery of reproductive 24-PHAS signal. Together, these results support Phasis as a biologically interpretable tool for large-scale discovery of plant DCL-dependent phasiRNA loci.

Introduction

Phased small interfering RNAs (phasiRNAs) are secondary small RNAs (sRNAs) produced from precursor transcripts after a trigger cleavage event establishes a phased register. They are termed secondary because their biogenesis is triggered by an upstream miRNA-guided cleavage event rather than by direct processing of the original precursor alone. In plants, this process depends on conversion of the target RNA into double-stranded RNA and subsequent Dicer-like (DCL) processing, which generates regularly spaced 21- or 24-nt phasiRNAs (Allen et al. 2005; Johnson et al. 2009; Zhan and Meyers 2023). Here, “phased” refers to the register-dependent production of sRNAs at regular intervals from an initiating cleavage site through successive DCL processing, distinguishing phasiRNAs from other DCL products, such as miRNAs, that are processed by DCL enzymes but do not define phased loci. The genomic loci whose transcripts give rise to phasiRNAs are referred to as PHAS loci.

Plant PHAS loci are diverse in precursor type, developmental context, and evolutionary distribution. Classic Arabidopsis trans-acting small interfering RNA (tasiRNA) loci, commonly referred to as TAS loci, established the mechanistic link between miRNA-directed cleavage and phasiRNA production (Allen et al. 2005; Peragine et al. 2004; Vazquez et al. 2004) whereas later studies identified phasiRNAs from protein-coding transcripts, including resistance-gene and PPR families, as well as abundant reproductive 21- and 24-nt phasiRNAs in many angiosperms (Fei et al. 2013; Xia et al. 2019; Zhai et al. 2011^{, 2015}). Because PHAS loci arise from different precursor types and operate in different biological contexts, their discovery can reveal regulatory pathways associated with development, reproduction, and defense. Accurate annotation of these loci is also increasingly important for plant genome resources, which need to represent functional noncoding RNA loci alongside protein-coding genes. The Kronos wheat functional genomic resource provides one example, with genome-wide annotations of both microRNAs and phasiRNAs (Seong et al. 2026).

Computational PHAS-locus discovery is challenging because a locus may appear to produce many sRNAs in a sequencing dataset for reasons unrelated to phasiRNA biogenesis. Genuine PHAS loci must therefore be distinguished from other genomic regions or transcripts that accumulate sRNA reads but lack coherent DCL-dependent phased processing. A well-supported PHAS locus requires more than local and periodic sRNA abundance patterns: the reads should show the expected 21- or 24-nt size class, a coherent phased register, and support across multiple phased positions. We use “phasiRNA signal” to describe this coordinated local pattern, because abundant background sRNA populations, especially heterochromatic 24-nt siRNA loci, can show high abundance or apparent periodicity without representing genuine phasiRNA biogenesis (Polydore et al. 2018). These features favor approaches that combine statistical evidence for phased accumulation with locus-level properties rather than relying on a single score or fixed parameter alone.

Phased sRNA patterns also occur outside plants, most notably in animal piRNA pathways, but these systems differ mechanistically from plant DCL-dependent phasiRNA biogenesis (Anleu Gil and Meyers 2026). Drosophila phased piRNAs are generated through Zucchini-dependent processing (Mohn et al. 2015; Han et al. 2015) and have broader length distributions than plant DCL-dependent phasiRNAs (Anleu Gil and Meyers 2026). These differences provide a useful boundary for evaluating whether a plant-focused PHAS-identification tool captures DCL-dependent phasiRNA architecture rather than generic or apparent periodic sRNA accumulation.

A set of command-line tools has been developed for de novo PHAS-locus discovery, including PhaseTank, PhasiHunter, DIGITAL, and earlier versions of ShortStack (Guo et al. 2015; Bu et al. 2023; Feng et al. 2023; Axtell 2013). However, robust inference across heterogeneous plant datasets remains difficult because PHAS classes differ in abundance, complexity, strand bias, phase length, and background sRNA composition. This creates a need for software that is reproducible across datasets while still allowing evidence thresholds and interpretation to reflect biological context. We previously developed a computational tool called PHASIS for de novo discovery and characterization of phasiRNA-generating loci (Kakrana et al. 2017). Here, we describe Phasis version 2, hereafter Phasis, a substantially redesigned and expanded successor focused on robust plant PHAS-locus discovery and interpretation from sRNA sequencing data.

Phasis identifies candidate PHAS loci by evaluating whether sRNAs associated with a genomic region show features expected from DCL-dependent phased processing, including enrichment for a dominant sRNA size class, regularly spaced reads, sufficient abundance, and evidence that the signal is unlikely to arise from random sRNA accumulation. Candidate loci are then evaluated by a Register-Resolved Locus Interpretation Layer (RRL). Here, “register-resolved” means that Phasis does not treat a candidate locus only as a broad interval with a single score; instead, it inspects phased positions within the locus to identify the dominant register, define the main phased unit, and distinguish coherent phased structure from secondary, overlapping, or unsupported signals. Using benchmark datasets across multiple plant systems, we show that Phasis recovers validated 21- and 24-PHAS loci with strong false-positive control, performs well relative to existing tools, and reports biologically interpretable locus structures. We further use maize dcl5 mutant libraries to validate pathway-specific 24-PHAS recovery and Drosophila piRNA-producing libraries as a boundary test for the plant-focused scope of the software.

Results

Overview of the Phasis Software Workflow

We developed Phasis to identify plant PHAS loci from sRNA sequencing (sRNA-seq) data by evaluating whether sRNAs originating within genomic or transcript-derived clusters show the size, abundance, and positional organization expected from DCL-dependent phased processing. The workflow has two stages (Fig. 1A). First, Phasis performs Initial PHAS Candidate Detection, in which sRNA-producing clusters are identified and scored using computed evidence features such as size-class enrichment, local abundance, strand bias, and phased-register support. Second, candidate loci are examined by a Register-Resolved Locus Interpretation Layer (RRL), which summarizes how the phasiRNA signal is organized within each candidate locus. Library-resolved summaries of PHAS detection and abundance can then be used to compare phased loci across samples or biological conditions, providing a genome-wide view of candidate PHAS signal across libraries (Fig. 1B).

Figure 1.

Phasis workflow and representative register-resolved outputs.

(A) Phasis analyzes sRNA libraries together with a genome or transcriptome reference to identify candidate PHAS loci. After library preparation, reference indexing, sRNA mapping, cluster discovery, and locus-level evidence calculation, initial candidate PHAS loci are passed to the Register-Resolved Locus Interpretation Layer (RRL). RRL selects a register anchor based on the highest phased-score position (HPSP), defines the main phased unit, evaluates coherent register extension, tests opposite-strand partners, assigns secondary or alternative windows, and interprets candidate locus architecture. Outputs include high-confidence PHAS calls, PHAS locus heat maps, GFF3 (a standard genome annotation format that describes genomic features) and tabular files, diagnostic plots, and phasiRNA annotations.

(B) Representative genome-wide Phasis heat map showing candidate PHAS signal across chromosomes and sRNA libraries. Rows correspond to developmental-stage libraries, columns correspond to genomic positions ordered by chromosome, and colored marks indicate sRNA-producing intervals summarized across genomic position. Red indicates candidate PHAS loci, dark blue indicates sRNA-producing loci not called as PHAS, and light blue indicates weak or no detected signal.

(C) Representative locus-level diagnostic output. The top plot shows local sRNA abundance across genomic position, with reads displayed above and below the midline according to mapped strand. This strand-separated display reflects the double-stranded precursor context expected for plant phasiRNA biogenesis and helps distinguish coherent phased signal from unsupported accumulation. Colors mark sRNA size classes and shaded vertical bands mark register-resolved phased windows. The bottom plot shows Howell-score support across the same genomic interval, again separated by strand, illustrating where phased support is concentrated relative to the interpreted register-resolved windows. A more detailed explanation of the color and symbol conventions used in these diagnostic plots, together with representative high-confidence, ambiguous, and unsupported candidate loci, is provided in Fig. S1.

During Initial PHAS Candidate Detection, Phasis evaluates evidence for phasiRNA accumulation using multiple locus-level features. The Phasis score summarizes statistical support for phased accumulation, whereas the Howell score provides an adapted register-level phased-score metric based on the formula described by Howell et al. (2007). In Phasis, the Howell score allows strand-specific positional variance of plus or minus one nucleotide at the 5′ end of the expected phase register, an implementation choice intended to tolerate limited positional variation while preserving register-level evidence for DCL-dependent phased processing. Detailed definitions of these features and the full Phasis score calculation are provided in the Supplemental Methods.

The RRL adds a post-detection interpretation step. Rather than treating each candidate as a single undifferentiated interval, this layer examines the local phased pattern and summarizes whether the signal forms a clear main phased unit, includes opposite-strand support, or contains additional supported phased windows. Phasis also reports locus-level diagnostic plots, inspired by the web-based MPSS/Plant small RNA database phasing-analysis viewer described by Nakano et al. (2020), that display local sRNA abundance, strand organization, size-class composition, register-resolved windows, and Howell-score support across the candidate interval (Fig. 1C). A detailed guide to these diagnostic plots, including color and symbol conventions and representative high-confidence, ambiguous, and unsupported candidate loci, is provided in Fig. S1. Together, these outputs allow Phasis to report both candidate-level evidence and a biologically interpretable model of phasiRNA organization within each locus.

Phasis Improves the Precision–Recall Balance Across Plant PHAS Benchmarks

We first evaluated whether Phasis could recover validated or annotated PHAS loci across diverse plant datasets. The benchmark set included maize and coffee 21-PHAS and 24-PHAS references, curated reproductive 24-PHAS references from petunia and columbine, and annotated Arabidopsis TAS loci. These datasets were selected to represent multiple discovery contexts, including reproductive 24-nt phasiRNA-producing loci, protein-coding or noncoding phased precursors, and a negative 24-PHAS context in Arabidopsis, which lacks the miR2275–DCL5 reproductive 24-nt phasiRNA pathway. This range of datasets also tested performance in 24-nt-rich plant backgrounds, where abundant 24-nt siRNA populations can lead to false-positive 24-PHAS annotation (Polydore et al. 2018). Here, individual-library evidence refers to loci detected in one or more constituent libraries, whereas pooled-library evidence refers to loci detected after libraries were combined before candidate detection.

Across these plant benchmarks, Phasis recovered expected PHAS loci with a strong balance between reference recovery and false-positive control. In petunia flower bud libraries (Xia et al. 2019) aligned to the reference genome described by Bombarely et al. (2016), pooled-library Phasis analysis recovered 367 of 440 curated 24-PHAS reference loci with 11 false-positive calls, corresponding to a precision of 0.97, recall of 0.83, and F1 score of 0.90. The F1 score summarizes the balance between precision and recall as their harmonic mean. These results support robust 24-PHAS identification despite the alternative phased-register architecture previously described in petunia (Xia et al. 2019). In columbine, using the genome assembly described by Filiault et al. (2018), Phasis recovered 466 of 576 curated 24-PHAS reference loci in pooled-library analysis, with a precision of 0.92, recall of 0.81, and F1 score of 0.87.

Arabidopsis provided a complementary validation context. Rather than benchmarking reproductive PHAS recovery against a positive reference, we tested whether Phasis recovered well-characterized TAS-associated 21-PHAS loci while avoiding unsupported interpretation of 24-PHAS candidates in a species lacking the miR2275–DCL5 reproductive 24-nt phasiRNA pathway. In sRNA libraries from 12-day-old Arabidopsis seedlings (Li et al. 2016), pooled-library analysis recovered six of eight annotated TAS loci. AtTAS3b and AtTAS3c were not recovered by either evidence mode, possibly reflecting limited local phasiRNA abundance in bulk seedling libraries given the spatially restricted activity of the miR390–TAS3 pathway during lateral-root development (Marin et al. 2010). Phasis also produced no high-confidence individual-library 24-PHAS calls in Arabidopsis, consistent with the absence of the miR2275–DCL5 reproductive 24-nt phasiRNA pathway in this species.

We next compared pooled-library Phasis outputs with ShortStack v3.8.5 and PhaseTank v1.0 using matched benchmark datasets and the same curated reference sets (Fig. 2A–C; Supplemental Table S1). These comparisons were restricted to combined-evidence outputs because the comparator tools do not provide the same individual-library evidence mode as Phasis. Across these matched analyses, Phasis generally provided the strongest balance between reference recovery and false-positive control, as summarized by F1 score (Fig. 2A), while the true-positive and false-positive call counts are shown in Fig. 2B. In petunia 24-PHAS analysis, Phasis achieved an F1 score of 0.90, compared with 0.64 for ShortStack and 0.58 for PhaseTank. In columbine 24-PHAS analysis, Phasis again had the highest F1 score, 0.87, compared with 0.41 for ShortStack and 0.70 for PhaseTank.

Figure 2.

Pooled-library benchmark comparison of Phasis, ShortStack, and PhaseTank across plant PHAS datasets.

(A) F1 scores calculated from call-level precision and reference-locus recall after PhasMatch comparison against curated reference sets for petunia 24-PHAS, columbine 24-PHAS, coffee 21-PHAS and 24-PHAS, and maize wild-type anthers 21-PHAS and 24-PHAS benchmarks. To enable direct comparison, this benchmark shows only pooled-library Phasis results, because ShortStack and PhaseTank were evaluated using pooled-library input and do not provide the individual-library evidence mode used by Phasis. Software colors are consistent across all panels and figures.

(B) Threshold-positive call counts for each pooled-library benchmark and tool, split into true-positive and false-positive calls. True positives are calls matched to curated reference loci, whereas false positives are threshold-positive calls not matched to the corresponding reference set.

(C) Call-level precision versus reference-locus recall for each pooled-library benchmark and tool. Circles indicate 21-PHAS analyses and triangles indicate 24-PHAS analyses. Point colors indicate software as in panel A and are used consistently across figures. The dashed diagonal marks equal precision and recall; points closer to the upper-right indicate stronger joint call-level precision and reference-locus recovery.

The strongest performance differences were observed in coffee. For coffee 21-PHAS loci, Phasis achieved an F1 score of 0.93, compared with 0.47 for ShortStack and 0.10 for PhaseTank. For coffee 24-PHAS loci, Phasis achieved an F1 score of 0.95, whereas ShortStack and PhaseTank had F1 scores of 0.13 and 0, respectively. In maize 21-PHAS analysis, Phasis also had the highest F1 score, 0.92, followed by PhaseTank at 0.79 and ShortStack at 0.75. Maize 24-PHAS remained the main exception in the pooled-library comparator benchmark, where Phasis and PhaseTank had similar F1 scores, 0.65 and 0.64, respectively, while ShortStack was lower at 0.45.

The precision–recall relationship was consistent with these benchmark results: across datasets, Phasis generally showed high precision and high recall, whereas comparator outputs more often showed reduced recall or lower precision depending on the dataset (Fig. 2C). Together, these plant benchmarks show that Phasis generally improved the balance between reference recovery and false-positive control across diverse PHAS discovery contexts. The maize 24-PHAS comparator result also indicates that pooled-library benchmarking captures only one operating mode of Phasis. We therefore separately evaluated how evidence mode affects recovery and precision.

Evidence Mode Affects PHAS Recovery and Precision

We next evaluated how individual-library and pooled-library evidence affected PHAS recovery under the final feature-threshold-enabled and RRL configuration. Across maize and coffee, pooled-library evidence generally increased reference-locus recovery, but the magnitude and precision effect depended strongly on dataset and phase class (Fig. 3A; Supplemental Table S2). In maize, pooled-library evidence modestly increased recall for both 21-PHAS loci, from 0.82 to 0.88, and 24-PHAS loci, from 0.82 to 0.86. In coffee, the effect was much larger, increasing recall from 0.36 to 0.88 for 21-PHAS loci and from 0.09 to 0.92 for 24-PHAS loci. These gains were especially clear for coffee 24-PHAS, where pooled-library evidence recovered 459 of 497 reference loci with a precision of 0.96 and F1 score of 0.95. By contrast, maize 24-PHAS pooled-library analysis increased recall but reduced precision, showing that pooled evidence can improve recovery while altering the precision–recall balance (Fig. 3A; Supplemental Table S2).

Figure 3.

Evidence mode and register-resolved interpretation affect PHAS recovery and false-positive control.

(A) Call-level precision, reference-locus recall, and F1 score for individual-library and pooled-library Phasis analyses under the default configuration. Results are shown for maize and coffee 21-PHAS and 24-PHAS benchmarks.

(B) Reference-locus overlap between individual-library and pooled-library evidence under the same feature-threshold-enabled and RRL-enabled configuration. Gray segments indicate reference loci recovered by both evidence modes. Colored segments indicate loci recovered only by one evidence mode, with blue denoting 21-PHAS benchmarks and gold denoting 24-PHAS benchmarks. Numbers indicate recovered reference loci in each segment. Pooled-library evidence increased total recovery in several contexts but was not a strict superset of individual-library detection.

(C) Effect of RRL on false-positive control under feature-threshold-disabled detection settings. Left, false-positive call counts with and without RRL for maize 21-PHAS individual-library analysis and coffee 24-PHAS pooled-library analysis. Right, corresponding call-level precision values. RRL reduced false-positive inflation and improved call-level precision by requiring coherent register-resolved phased architecture.

Overlap analysis using the reference loci as the baseline showed that detection from pooled libraries was not a strict superset of detection from individual libraries (Fig. 3B). In maize, pooling libraries recovered more reference loci overall, but analyses of individual libraries still detected 46 21-PHAS and 14 24-PHAS loci that were missed after pooling. The same pattern was observed in coffee: pooling libraries strongly increased recovery, but analyses of individual libraries still detected 22 21-PHAS and 16 24-PHAS loci that were not recovered from the pooled data. Thus, pooling changed the evidence profile rather than simply adding statistical power to detection from individual libraries.

The additional plant validation datasets extended this pattern beyond the maize and coffee benchmarks (Supplemental Table S3). In petunia, pooling libraries substantially increased recovery of curated reproductive 24-PHAS loci while retaining high precision. In columbine, pooling also increased recovery of reference loci, but it produced a larger number of false positives and reduced precision relative to analyses of individual libraries. In Arabidopsis, pooling rescued one additional annotated TAS locus, AtTAS4, that was not recovered from the individual libraries. These results indicate that the evidence mode should be selected according to the dataset structure and biological question: analyses of individual libraries preserve sample-specific signals, whereas pooling can recover weak or distributed signals but may also change the false-positive profile.

Register-Resolved Interpretation Controls False-Positive Inflation

Because pooled-library evidence could increase recovery while also altering the false-positive profile, we next tested how much the RRL contributed to false-positive control. Maize and coffee 21-PHAS and 24-PHAS benchmarks were evaluated under matched configurations that varied feature-threshold status, RRL status, and evidence mode. Feature-threshold-disabled analyses represented a deliberately permissive candidate-detection setting, allowing us to test whether RRL could reduce unsupported calls after broad initial detection. The full benchmark, including all feature-threshold-enabled/disabled, RRL present/absent, individual-library, and pooled-library configurations, is provided in Supplemental Table S2.

Across the full benchmark, RRL had its strongest effect under permissive settings in which feature thresholds were disabled. In maize 21-PHAS individual-library analysis, feature-threshold-disabled detection without RRL produced 1,258 false-positive calls and a call-level precision of 0.34, whereas adding RRL reduced false-positive calls to 26 and increased call-level precision to 0.95, with only a small change in reference-locus recall. The same pattern was especially clear in 24-PHAS analyses with pooled-library evidence, where the number of unsupported calls was highest. For example, in coffee 24-PHAS pooled-library analysis, feature-threshold-disabled detection without RRL produced 15,574 false-positive calls and a call-level precision of 0.02; adding RRL reduced false-positive calls to 921 and increased call-level precision to 0.27 while retaining high reference-locus recall, illustrating the false-positive-control effect summarized in Fig. 3C.

These results show that RRL reduces false-positive inflation when initial candidate detection is intentionally broad. By assessing phased-register consistency rather than relying solely on signal abundance, RRL evaluates whether candidate loci contain coherent register-resolved phased architecture, allowing Phasis to separate better-supported PHAS calls from unsupported or ambiguous sRNA-producing regions. We next asked whether these high-confidence calls also reflected expected pathway biology in maize dcl5 mutant libraries.

dcl5 Mutant Benchmarking Supports DCL5-Dependent 24-PHAS Recovery

We next used maize dcl5 mutant anther libraries to test whether Phasis recovery of 24-PHAS loci reflected the expected DCL5-dependent biology of reproductive 24-nt phasiRNAs. The analysis compared W23 2.0 mm anther libraries with stage-matched homozygous dcl5 mutant 2.0 mm anther libraries, corresponding to the meiotic window in which 24-nt reproductive phasiRNAs are expected to accumulate (Teng et al. 2020).

Under the fixed-threshold and RRL configuration, homozygous dcl5 anther libraries showed strong depletion of maize 24-PHAS recovery relative to W23 libraries (Fig. 4A). This reduction was observed for both the full maize 24-PHAS reference set and the meiotic-enriched 24-PHAS reference subset, consistent with the expected dependence of reproductive 24-nt phasiRNA accumulation on DCL5.

Figure 4.

dcl5 mutant analysis in maize supports pathway-specific recovery of reproductive 24-PHAS loci.

(A) Reference-locus recall in W23 2.0-mm anther libraries and stage-matched homozygous dcl5 mutant 2.0-mm anther libraries under the default configuration.

(B) CP20M-normalized abundance distributions for called 24-PHAS loci and reference-overlapping non-called loci in W23 and homozygous dcl5 anther libraries. Raw CP20M values are displayed on a log-scaled y-axis for visualization. W23 called loci are shown in gold, dcl5 called loci in orange, and dcl5 reference-overlapping non-called loci as gray points over yellow box plots. N indicates W23 called loci, C indicates dcl5 called loci, and R indicates dcl5 reference-overlapping non-called loci.

(C) Phasis-score versus Howell-score diagnostic plot for W23 24-nt candidate filtering. Dashed lines mark the fixed-threshold criteria used for high-confidence 24-PHAS calling. W23 called 24-PHAS loci are concentrated in the high Phasis-score and high Howell-score region, whereas filtered false positives and non-PHAS background loci show insufficient or discordant evidence.

The abundance profiles provided a complementary view of this depletion. W23 libraries contained many high-abundance 24-PHAS calls, whereas homozygous dcl5 anther libraries contained few high-confidence 24-PHAS calls (Fig. 4B). In the dcl5–1 homozygous 2.0-mm anther library, only three loci were still called as 24-PHAS and showed higher median CP20M abundance than reference-overlapping loci that remained detectable but did not pass high-confidence calling criteria. In the dcl5-2 homozygous 2.0-mm anther library, no 24-PHAS loci passed these criteria. Together, these abundance patterns support strong depletion of the 24-PHAS signal in homozygous dcl5 anthers, with only a small number of high-abundance reference-overlapping loci remaining detectable as high-confidence calls.

Score-space analysis further separated supported 24-PHAS calls from threshold-conflict candidates and non-PHAS background clusters (Fig. 4C). Validated 24-PHAS calls were concentrated in the high Phasis-score/high Howell-score region, whereas many filtered candidates showed discordant evidence, including high Howell score with low Phasis score or high Phasis score with insufficient Howell-score support. Together, the reference-recovery, abundance, and score-space analyses indicate that Phasis detects a DCL5-dependent 24-PHAS signal rather than simply calling abundant 24-nt sRNA clusters.

Drosophila Scan Defines a Mechanistic Boundary for Phasis

We applied Phasis to sRNA libraries from Drosophila melanogaster ovarian tissues, a species and organ known to produce phased piRNAs through Zucchini-dependent processing (Han et al. 2015; Mohn et al. 2015). Because piRNAs span a broader size range than plant DCL-dependent phasiRNAs, we scanned phase periodicities from 23 to 30 nt. Across this scan, Phasis reported only three candidates: one 25-nt candidate and two 26-nt candidates.

Manual inspection and coordinate-level annotation based on the FlyBase annotation release 6.54 (Öztürk-Çolak et al. 2024) indicated that these calls did not represent convincing plant-like PHAS loci. Their apparent periodicity was supported by limited local signal rather than by distributed register-wide phased support across a coherent locus. One candidate occurred in a repeat-rich region containing a transposable-element insertion and DNAREP1_DM Helitron sequence, another contained multiple simple repeats, and the third was dominated by a single 25-nt sequence without broader phased support.

These results define a mechanistic boundary for Phasis. Mechanistically distinct sRNA populations, including piRNA-sized or repeat-associated signals, can occasionally produce fixed-spacing patterns that satisfy automated plant-oriented criteria. However, these candidates did not show the architecture expected for plant DCL-dependent PHAS loci when evaluated with locus plots and genomic annotations. Phasis is therefore intended for plant DCL-dependent phasiRNA discovery rather than for general detection of phased or periodic animal sRNA production.

Discussion

Phasis was designed to support the conservative and biologically interpretable discovery of plant PHAS loci from sRNA-seq datasets. The results show that combining Initial PHAS Candidate Detection with register-resolved interpretation improves the separation of high-confidence PHAS calls from ambiguous or unsupported sRNA-producing regions. Across plant benchmarks, Phasis recovered validated 21- and 24-PHAS loci with a strong balance between reference recovery and false-positive control, and comparator analyses showed improved performance relative to existing tools in most benchmark contexts.

The benchmark results also show that evidence mode affects PHAS recovery. Pooled-library analysis often recovered additional validated loci, especially when phasiRNA signal was weak or distributed across libraries, but it was not a strict replacement for individual-library analysis. Individual-library evidence preserved sample-specific signals and recovered some loci that were missed after pooling. Thus, evidence mode should be selected according to dataset structure and biological question rather than assuming that one approach is always more sensitive than the other.

The RRL and dcl5 analyses support the importance of evaluating PHAS candidates beyond abundance or apparent periodicity alone. RRL reduced false-positive inflation by requiring coherent register-resolved phased architecture, while the maize dcl5 comparison showed strong depletion of 24-PHAS recovery in homozygous mutant anther libraries. Together, these analyses indicate that Phasis detects pathway-consistent 24-PHAS signal.

A practical consideration is that conservative, register-resolved analysis can require substantial computational resources, particularly for samples enriched in 24-nt sRNAs analyzed in 24-PHAS mode. As a general guideline, we recommend allocating approximately 4 GB of RAM per library, although highly abundant or deeply mapped datasets may require more. In some 24-nt-enriched datasets, memory use in intermediate stages approached approximately 1.14 GiB per million mapped reads, and users should therefore adjust memory allocation and expected runtimes according to library size, mapping depth, and phase mode.

The Drosophila scan clarifies the intended scope of the software. Mechanistically distinct sRNA populations, including animal piRNAs and repeat-associated signals, can occasionally produce fixed-spacing patterns that satisfy automated plant-oriented criteria. These candidates require biological and annotation-based interpretation and should not be treated as evidence that Phasis has the capacity to be a general detector of animal phased sRNA pathways. Phasis is therefore best used for plant DCL-dependent phasiRNA discovery, where periodicity can be interpreted together with size class, abundance structure, strand/register support, and known pathway biology.

Together, these results support Phasis as a tool for register-resolved discovery and interpretation of plant PHAS loci across diverse sRNA-seq datasets. To support manual review and biological interpretation, Phasis also generates integrated locus-level diagnostic plots that summarize local sRNA abundance, size-class composition, strand organization, phased-register support, and Howell-score evidence. These visual summaries provide a practical check of candidate architecture and help distinguish coherent PHAS loci from overlapping, weakly supported, or ambiguous sRNA-producing regions.

Materials and Methods

Small RNA Datasets and Reference Genomes

Small RNA sequencing datasets were selected to represent multiple plant PHAS discovery contexts and one non-plant boundary case. Maize and coffee datasets were used for the main 21-PHAS and 24-PHAS benchmarks (Zhai et al. 2015; Cherubino Ribeiro et al. 2024); petunia and columbine were used as additional reproductive 24-PHAS validation datasets (Xia et al. 2019); Arabidopsis seedlings were used for annotated TAS-associated 21-PHAS recovery and as a negative 24-PHAS context (Li et al. 2016); and Drosophila ovarian libraries were used as a mechanistically distinct piRNA-producing system (Han et al. 2015).

For each species, sRNA reads were analyzed against the genome assembly used for the corresponding reference set and benchmark. Arabidopsis TAS validation was performed using Araport11 TAS gene annotations. Dataset accessions and sample metadata are provided in Supplemental Table S4. Curated PHAS locus annotations and coordinates used for validation are provided in Supplemental Table S5. Representative analysis workflows are provided in Supplemental File S1.

Phasis Analyses and Evidence Modes

Phasis analyses were performed with phase length set according to the PHAS class or small-RNA pathway being evaluated. Plant analyses focused on 21-nt and 24-nt phase lengths, corresponding to the plant phasiRNA classes analyzed in this study. Drosophila analyses used a phase-length scan from 23 to 30 nt to encompass the broader size range of ovarian piRNAs. Unless otherwise indicated, Phasis calls discussed as high-confidence PHAS loci refer to outputs passing the fixed-threshold and register-resolved interpretation criteria used for the corresponding benchmark.

Two evidence modes were evaluated where appropriate. In individual-library mode, PHAS calling was performed separately for each sRNA library, preserving library-specific phasiRNA accumulation, abundance, and register structure. In pooled-library mode, libraries belonging to the same dataset were analyzed as combined evidence before candidate detection, allowing weak signals distributed across libraries to contribute to the same locus. These modes were interpreted as complementary rather than hierarchical alternatives. When both evidence modes were available, overlap between recovered reference loci was evaluated to determine whether pooled-library detection acted as a strict superset of individual-library detection or provided a distinct evidence profile.

Reference Sets and PhasMatch Benchmarking

Species-specific reference coordinates were used to evaluate Phasis performance for each benchmarked species and phase class. These included curated maize, coffee, petunia, and columbine PHAS reference sets and Araport11 TAS gene coordinates for Arabidopsis. Arabidopsis 24-PHAS outputs were interpreted as a negative biological context because this species lacks the miR2275–DCL5 reproductive 24-nt phasiRNA pathway.

Phasis calls were compared with reference loci using the Phasis supporting tool PhasMatch. Each PHAS call was treated as a coordinate-defined call. Calls that overlapped a curated reference locus were counted as true-positive calls, whereas calls without reference support were counted as false-positive calls. Reference loci overlapped by one or more calls were counted as recovered reference loci, whereas reference loci not overlapped by any call were counted as unrecovered reference loci.

These outputs were used to calculate call-level precision, reference-locus recall, and F1 score. Precision was evaluated at the call level, whereas recall was evaluated at the reference-locus level. For analyses comparing individual-library and pooled-library evidence, reference-locus recovery was also summarized as the overlap between reference loci recovered in both evidence modes, recovered only in individual-library analyses, recovered only in pooled-library analyses, or not recovered by either mode. This reference-side overlap was used to determine whether pooled-library evidence behaved as a strict superset of individual-library evidence or provided complementary support.

dcl5 Mutant Validation and CP20M Abundance Analysis

Maize dcl5 mutant libraries were used to evaluate whether Phasis recovery of 24-PHAS loci reflected the expected DCL5-dependent biology of reproductive 24-nt phasiRNAs. The analysis compared W23 2.0-mm anther libraries with stage-matched homozygous dcl5 mutant 2.0-mm anther libraries. W23 and dcl5 libraries were analyzed using the same fixed-threshold and register-resolved interpretation criteria applied in the maize benchmark analyses.

Reference recovery was evaluated against the full maize 24-PHAS reference set and against a meiotic-enriched 24-PHAS subset. Maize 21-PHAS reference recovery was evaluated in the same dcl5 libraries as an internal control for general 21-nt phasiRNA detection. Recovery of reference loci was assessed using PhasMatch.

For abundance comparisons, 24-PHAS calls and reference-overlapping non-called loci were quantified using total locus abundance reported by Phasis. Abundance was normalized as CP20M by dividing total locus abundance by the corresponding sum of mapped reads for a given library and multiplying by 20,000,000. In dcl5 libraries, called 24-PHAS loci were compared with reference loci that retained detectable abundance but did not pass the high-confidence 24-PHAS calling criteria.

Candidate loci were also visualized in Phasis-score and Howell-score space. Validated 24-PHAS calls, filtered false-positive candidates, dcl5-associated candidates, and non-PHAS background loci were plotted relative to the fixed-threshold criteria used for 24-PHAS calling. This diagnostic was used to evaluate whether candidate loci showed concordant statistical and register-level support or instead represented threshold-conflict candidates with incomplete evidence for high-confidence 24-PHAS annotation. CP20M-normalized abundance distributions were summarized descriptively using counts and median CP20M values for called 24-PHAS loci and reference-overlapping non-called loci.

Additional Plant Validation Datasets

For petunia and columbine, individual-library and pooled-library Phasis outputs were compared with the corresponding curated reproductive 24-PHAS reference sets using PhasMatch. Performance was summarized using call-level precision, reference-locus recall, F1 score, and reference-side overlap between evidence modes.

For Arabidopsis, individual-library and pooled-library Phasis outputs were compared with Araport11 TAS annotations to evaluate recovery of annotated TAS-associated 21-PHAS loci. Pooled Arabidopsis 24-PHAS candidates were manually inspected against Araport11 gene and transposable-element annotations. Candidates that lacked convincing phased-register structure or overlapped annotated transposable-element genes were classified as unsupported rather than as bona fide reproductive 24-PHAS loci.

Drosophila Phase-Length Scan and Annotation Audit

Drosophila ovarian sRNA libraries were analyzed as a non-plant boundary test because animal piRNAs can show phased production through Zucchini-dependent processing, but this pathway differs mechanistically from plant DCL-dependent phasiRNA biogenesis. The analysis used ovarian libraries from Han et al. (2015) and the corresponding FlyBase release 6.54 genome assembly. Because piRNAs have a broader size distribution than plant phasiRNAs, Phasis was run across phase lengths from 23 to 30 nt.

Automated calls from the Drosophila phase-length scan were manually inspected to determine whether they showed plant-like PHAS architecture or instead reflected local abundance artifacts, repeat-associated sRNAs, or piRNA-sized signals that could mimic fixed-register periodicity. For the three candidates reported by Phasis, locus plots were examined for coherent locus-level organization.

Candidate coordinates were then compared with FlyBase release 6.54 gene, transposable-element, repeat, and other genomic annotations. Calls were interpreted as unsupported when their apparent periodicity could be explained by repeat-rich or transposable-element-associated sequence context, concentrated local abundance, piRNA-associated sequence features such as 1U status where relevant, or other patterns inconsistent with plant-like PHAS architecture.

Comparison with Existing Tools in Matched Pooled-Library Analyses

To evaluate Phasis relative to existing PHAS-discovery tools, we performed matched pooled-library benchmarks using Phasis, ShortStack v3.8.5, and PhaseTank v1.0 outputs generated from the same sRNA-library sets and evaluated against the same curated reference sets. Direct tool-to-tool comparisons were restricted to pooled-library analyses because the comparator tools do not provide the same individual-library evidence mode as Phasis. Individual-library Phasis results were therefore not used for direct comparator benchmarking.

For each species and phase class, PHAS calls from each tool were compared with the corresponding curated reference set using PhasMatch. Benchmarking was summarized using call-level and reference-locus metrics. Using the metric definitions described above, call-level precision quantified the proportion of calls supported by curated reference loci, whereas reference-locus recall quantified recovery of curated loci. F1 score summarized the balance between call-level precision and reference-locus recall. For ShortStack, positive calls were restricted to records with PhaseScore ≥ 12.5, and for PhaseTank, positive calls were restricted to records with Phased_Score ≥ 12.5. Although these metrics are named differently across software packages, all are derived from the phased-locus scoring framework originally described by Howell et al. (2007). Each tool implements its own adaptation of that framework, resulting in software-specific score calculations and nomenclature. These threshold values were selected to match the Howell-score cutoff used for Phasis positive-call filtering before benchmarking. The main comparator benchmark table is provided as Supplemental Table S1.

Data Availability

The sequencing datasets analyzed in this study are available from Gene Expression Omnibus (GEO) or the Sequence Read Archive (SRA). Benchmark summary tables, sequencing-library metadata and accession identifiers, curated PHAS locus annotations and coordinates, and representative analysis workflows are provided in Supplemental Tables S1–S5 and Supplemental File S1.

Code Availability

Phasis is an open-source Python software package available from the project GitHub repository https://github.com/blakemeyerslab/Phasis under the OSI Artistic License 2.0. The archived, versioned release corresponding to the implementation used in this study is available at Zenodo (DOI: 10.5281/zenodo.20435283). The software was tested with Python 3.12 on Unix-like operating systems, including Linux and macOS. Installation is provided via pip from the repository, which installs the Phasis command-line interface; HISAT2 and SAMtools must be available on the user’s PATH. A detailed user manual, including installation instructions, system requirements, command-line options, and example analysis workflows, is provided in the repository.