SpikeForge–FusionSieve Fusion Recovery Report

Abstract

Summary

The primary hypothesis was that FusionSieve would recover the correct fusion partner pair, orientation, and both breakpoints within 5 base pairs in each modified BAM, while the matched original BAM remained negative for that target. This hypothesis was supported in all seven controls. FusionSieve recovered every junction within 2 bp, assigned every call a score of 1.0, and applied no hard filters. All calls remained at REVIEW, as expected, because automatic reporting was disabled and no validated scorer or compatible panel of normals was used.

SpikeForge added 1,681 BAM records in total. Its record-level verifier found zero removed or unexpected records, FusionSieve independently read the complete record count for all seven outputs, and all seven generated BAI files passed a separate structural traversal. However, the post-hoc realism hypothesis was not supported: nearby real alignments carried AS and XS tags, while the synthetic alignments did not. The result therefore supports functional software interoperability only—not clinical sensitivity, specificity, limit of detection, or treatment use.

7 / 7

Target junctions recovered

Five EML4–ALK and two CD74–ROS1 controls.

0 / 7

Targets in matched baselines

No exact target candidate was present before spiking.

≤ 2 bp

Breakpoint error

Inside the predeclared ±5 bp recovery window.

0

Hard-filtered controls

All seven recovered calls had an empty hard-filter list.

Study question

Hypotheses and endpoints

Primary hypothesis

A truth-locked FusionSieve run, reconciled only after calling, will recover the planted gene pair, strand orientation, and both breakpoints within ±5 bp in each modified BAM, with no target candidate in its matched baseline.

Secondary hypotheses

SpikeForge will add only declared synthetic records and preserve every original record.
FusionSieve will parse every complete output and retain independent split and spanning evidence.
The planted BAM alignments will be indistinguishable from nearby real alignments in the realism audit.

Predeclared interpretation of each endpoint
Endpoint	Criterion	Observed status
Caller recovery	Correct pair, orientation, and both positions within ±5 bp	Supported
Baseline specificity within this set	No exact target candidate in the matched original BAM	Supported
Record preservation	Zero removed or unexpected records	Supported
BAM-level realism	No field that separates planted from nearby real alignments	Not supported
Clinical performance	Requires assay-specific reference materials and a prospective design	Not tested

Experimental conduct

Methods

Select and lock backgrounds. Seven distinct duplicate-removed, coordinate-sorted Illumina hg19 BAMs were assigned anonymous labels C01–C07. Their sizes and SHA-256 identities were locked before modification; all seven source hashes were rechecked successfully after the experiment.
Lock fusion truth. C01–C05 received EML4–ALK and C06–C07 received CD74–ROS1. CD74 was selected as the ROS1 partner because both genes are present in the locked FusionSieve hg19 panel; other ROS1 partners are not represented by that panel.
Generate modified BAMs. SpikeForge 1.1.0rc2 requested 60 split-support source fragments and 20 spanning pairs per control, using seeds 20260801–20260807. Each fusion was placed in a different BAM.
Run the caller without retuning. FusionSieve 1.1.0.dev0 was run independently on each original and modified BAM with the same audited no-PoN configuration. The complete JSON candidate set, including rejected candidates, was retained.
Reconcile after calling. Target identity, strand, and breakpoint distance were compared with the locked truth only after both baseline and modified runs completed. Tier, filters, and evidence counts were recorded separately from the binary recovery endpoint.

Locked fusion definitions
Control group	5′ partner	3′ partner	Requested controls
C01–C05	EML4 · chr2:42,522,656 · +	ALK · chr2:29,446,394 · −	5 independent BAMs
C06–C07	CD74 · chr5:149,784,243 · −	ROS1 · chr6:117,645,578 · −	2 independent BAMs

Reference and provenance details

Assembly: GRCh37/hg19. FusionSieve panel FASTA SHA-256: b7a5a05d5cf5d649fd74713e52865b9a106e5a548af834c6ac8e400c5c9ff0ed. Panel metadata SHA-256: c32bfdace27a220b53652918cda0111ca0582dc00a87c966c53f6bf4a4fbff09. FusionSieve runtime-source SHA-256: 8d9ec7e47445917fda75422e9e0152ead8e0b9c08a34bd3703d21de80b0be35c.

The available BAMs had no reconstructable reference coverage at the selected EML4 breakpoint. SpikeForge therefore used a strict genomic-coordinate adapter over the exact hg19 panel FASTA; any sequence request not fully contained in one panel tile failed closed. At covered ALK, CD74, and ROS1 loci, panel and BAM-derived reference sequence agreed at every compared base. Sharing the panel sequence guarantees assembly compatibility but is also a study limitation because the generator and caller were not reference-resource independent.

Primary evidence

Results

01

All seven fusion controls were recovered

Each modified BAM gained exactly one target candidate and each matched baseline had none. The maximum absolute breakpoint discrepancy was 2 bp. Every target call had the correct orientation, an empty hard-filter list, and a score of 1.0.

Per-control caller recovery
Control	Fusion	Baseline target	Detected	Breakpoint error, bp	Tier	Unique split	Unique span	Conservative support
C01	EML4–ALK	0	Yes	−2 / +2	REVIEW	50	19	69
C02	EML4–ALK	0	Yes	−2 / +2	REVIEW	51	20	71
C03	EML4–ALK	0	Yes	−2 / +2	REVIEW	56	20	76
C04	EML4–ALK	0	Yes	−2 / +2	REVIEW	59	19	77
C05	EML4–ALK	0	Yes	−2 / +2	REVIEW	58	20	78
C06	CD74–ROS1	0	Yes	+2 / +2	REVIEW	54	15	69
C07	CD74–ROS1	0	Yes	+2 / +2	REVIEW	53	11	64
Total caller evidence						381	124	504

“Conservative support” is FusionSieve’s minimum of panel-signature and BAM-coordinate-family support when coordinate keys are complete. It is a support proxy, not a molecule count and not tumor allele fraction.

02

Generation and record accounting passed

SpikeForge planted every requested split-support source fragment. It placed 137 of 140 requested spanning pairs; C07 could accommodate 17 of 20 because the input library’s short-insert distribution made three requested spans physically impossible. Actual support, not requested support, was used for interpretation.

420 / 420split fragments planted

567split-aligned read ends added

137 / 140spanning pairs planted

1,681total BAM records added

0original records removed

0unexpected record changes

03

The outputs were readable across the two codebases

SpikeForge’s verifier passed all seven outputs. FusionSieve then performed a complete independent BAM parse and its record count matched SpikeForge’s declared output count in all seven cases. A separate standard-library BAI traversal validated the internal structure and virtual-offset bounds of all seven indexes.

Software and file-integrity checks
Check	Result	Meaning
Locked input SHA-256 recheck	7 / 7 passed	The selected source files matched the predeclared inputs.
SpikeForge record-level verifier	7 / 7 passed	No original record was removed or unexpectedly changed.
FusionSieve full BAM parse	7 / 7 matched	Independent caller record totals matched the generated BAM totals.
BAI structural traversal	7 / 7 passed	All index members, chunks, intervals, and virtual offsets were traversable.
SpikeForge automated checks	182 + 48 passed	Focused pytest and original registry suites passed; one focused test skipped.
FusionSieve automated checks	143 passed	One test skipped and eight parameterized subtests passed.

04

Important negative result

All seven BAM-level controls failed the realism audit

AS and XS alignment-score tags were present on all sampled nearby real reads and absent from the planted reads. This makes the synthetic records distinguishable and could cause a downstream step that requires these tags to remove them. SpikeForge did not copy scores from unrelated reads, because that would create false alignment evidence; without running an aligner it cannot calculate truthful AS/XS values.

The controls are therefore suitable for this functional caller challenge but should not be treated as blind-realistic BAM controls. FASTQ-level generation followed by the production aligner is the appropriate next experiment.

Read the completed FASTQ-path follow-up, including the original-data inventory, structural tag comparison, seven matched caller runs, and the remaining production-realignment requirement.

Interpretation

Conclusions

Functional interoperability was demonstrated. FusionSieve recovered all seven SpikeForge fusion controls against exact-target-negative matched baselines.
The breakpoint and evidence paths behaved consistently. All calls were correctly oriented, within 2 bp, and free of hard filters.
File preservation was supported. All outputs passed exact record accounting, full downstream parsing, and BAI structural traversal.
BAM-level realism was not demonstrated. Missing AS/XS tags make the planted alignments separable from the local background.
No clinical claim follows. This small, high-support computational challenge does not establish sensitivity, specificity, LoD, or patient benefit.

Boundaries of the evidence

Limitations

The design used seven selected BAMs and one fixed high-support condition per fusion, not a replicated support ladder.
The generator and caller shared the exact hg19 panel sequence; an independently sourced assembly reference was not used.
The selected EML4 breakpoint had no reconstructable background reference depth in these BAMs, producing an expected low-depth soft flag.
No compatible panel of normals or independently validated FusionSieve model was supplied; p/q values are prior-only and must not be interpreted clinically.
FusionSieve support fractions are evidence proxies, not tumor fraction, molecular VAF, or validated molecule counts.
BAM-level insertion bypasses extraction, library preparation, sequencing, alignment, and duplicate-processing effects.
Native samtools, Picard, and IGV were not available for this run. File checks used SpikeForge verification, FusionSieve full parsing, and an independent BAI structural traversal.
Raw BAMs and operational caller JSON were not published because they contain sample-level sequence data and processing metadata.

Next experiments

Recommended future work

1

Complete production realignment

The FASTQ-path caller experiment is complete. The remaining step is to run those molecules through the pinned production aligner so AS, XS, MAPQ, CIGAR, and split-alignment flags are generated together.

2

Build a support ladder

Repeat each fusion at multiple split/span levels, with replicated seeds and blinded backgrounds, to estimate computational recovery curves.

3

Use independent resources

Generate junction sequence from a separately verified hg19 FASTA and evaluate with a locked, assay-compatible panel of normals and calibrated model.

4

Add orthogonal validation

Inspect representative alignments in IGV and confirm selected controls with a second fusion caller and, where appropriate, laboratory reference material.

Reproducibility and handoff

Download the validation materials

The public kit contains the locked plan, fusion definitions, anonymized per-control results, provenance summaries, reconciliation scripts, and checksums. It deliberately excludes BAMs, BAM indexes, read names, raw caller JSON, local paths, and sample identifiers.

Fusion interoperability validation kit

ZIP archive · protocol, scripts, sanitized data, and integrity manifest

605efb47abdeef5ef1340c9ae56a7eea7c8b110218b525bbf510cb3c9bbcd74b

Download ZIP SHA-256 file

Per-control resultsCSV Sanitized summaryJSON

SpikeForge

Generate local research controls and review its separate software-validation report.

Open SpikeForge report and download

FusionSieve

Review the caller’s methods, aggregate validation, limitations, and software package.

Open FusionSieve report and download

Where are the modified BAMs? They remain in the private experiment folder on the originating computer. They are not part of the public download because BAMs can contain sensitive sequence data. Colleagues can use the public kit with their own approved BAMs and local copies of the two research tools.