Computational validation report

Full-reference FASTQ-to-BAM fusion replay

Seven matched controls were processed through pinned BWA-MEM alignment against whole hg19, followed by FusionSieve analysis of the resulting BAM files.

1

Introduction and summary

An earlier audit found that fusion reads inserted directly into BAM files lacked alignment scores such as AS and XS and used a different split-alignment structure from nearby BWA-MEM records. A subsequent targeted FASTQ experiment recovered all seven planted fusions, but it did not create whole-genome alignments. This study performed the missing step: paired FASTQs were aligned against the full hg19 reference with the recorded historical BWA-MEM version and -M behavior, sorted and indexed, and then analyzed as BAMs by FusionSieve.

Exactly one target-window candidate was found in each of seven spiked controls, while none was found in the seven matched backgrounds. All 1,106 planted primary read records mapped. BWA-MEM generated AS, XS, NM, MD, and read-group tags on every mapped planted primary record. The maximum absolute breakpoint error was 2 bp.

7 / 7Spiked controls with one target-window candidate
0 / 7Matched backgrounds with a target-window candidate
100%Mapped planted primary records with AS and XS
0 / 7Candidates automatically reportable as PASS/HIGH
Bottom line

The full-reference replay solved the alignment-realism defect identified in the direct BAM experiment: the aligner, rather than the spike-in program, calculated the coupled alignment fields. It also reproduced candidate recovery in all seven spiked controls. The calls remain REVIEW-tier research evidence and are not clinical or automatically reportable results.

2

Study question and hypotheses

Primary question

Can FASTQ-level fusion generation followed by whole-reference BWA-MEM alignment produce internally coherent synthetic BAM records while preserving matched-background fusion recovery?

Working hypotheses

  1. Alignment hypothesis: BWA-MEM will calculate alignment scores, edit-distance fields, clipping, flags, mate fields, and split-hit structure from sequence and reference.
  2. Recovery hypothesis: each spiked BAM will contain one FusionSieve candidate within the predeclared ±5 bp target window.
  3. Matched-background hypothesis: the corresponding unmodified BAM will contain no candidate in that target window.
  4. Separation hypothesis: candidate detection and automatic reportability will remain distinct endpoints.
These are computational validation hypotheses. They do not address patient outcomes, treatment selection, analytical sensitivity, clinical sensitivity, or clinical utility.

3

Experimental design

Seven post-processed BAM backgrounds were represented by neutral control numbers. Five controls received an EML4–ALK design and two received a CD74–ROS1 design. Each control was analyzed twice: once as an unmodified matched background and once after FASTQ-level spike-in. This produced 14 independently aligned and analyzed BAM datasets.

Table 1. Study design and locked endpoints.
ComponentDesignPurpose
Matched controls7 backgrounds and 7 spiked counterpartsSeparate spike-in recovery from pre-existing background calls
Fusion definitions5 EML4–ALK; 2 CD74–ROS1Test two genomic fusion architectures
Synthetic dose553 read pairs: 420 split fragments and 133 spanning pairsProvide both junction and mate-pair evidence
Alignment analyses14 whole-hg19 BWA-MEM runsLet the aligner calculate coupled BAM fields
Primary endpointCorrect genes, chromosomes, strands, and both breakpoints within ±5 bpMeasure target-window candidate recovery
Negative controlZero target-window candidates in each matched backgroundRequire a paired, background-specific comparison

4

Methods

  1. Reconstruct paired FASTQ. Complete primary read pairs were recovered from the surviving post-processed BAMs and returned to sequencer orientation. This is a controlled reconstruction, not recovery of the absent instrument FASTQs.
  2. Remove coordinate-order clustering. Because the source BAMs were coordinate sorted, paired records were deterministically reordered by a pair-preserving HMAC-SHA256 key. The procedure verified matching mates and an unchanged pair multiset. It cannot recreate the original flow-cell order.
  3. Generate fusion read pairs. SpikeForge created the two fusion designs at the FASTQ level. It did not write BAM alignment scores, flags, CIGAR strings, or mate placements.
  4. Align against full hg19. Pinned BWA-MEM 0.7.12-r1039 was run with the historical -M convention, eight threads, and the recorded single read-group structure. The 93-contig hg19 FASTA was locked by SHA-256 and its primary contig names and lengths matched the source BAM headers.
  5. Sort, index, and validate. Samtools 1.24 sorted and indexed each BAM. The workflow checked coordinate order, read-group consistency, index readability, synthetic-read alignment fields, and file hashes before accepting each output.
  6. Run matched FusionSieve analyses. FusionSieve analyzed all 14 BAMs under the same paired configurations used in the earlier FASTQ study. Candidate calls were compared with truth only after the runs completed.
  7. Aggregate without sequence data. The public report uses control-level and aggregate statistics only. Read names, source identifiers, local locations, BAMs, FASTQs, and reference indexes are excluded from public artifacts.

Why FASTQ plus BWA is better than assigning tags

Alignment fields are not independent decorations. AS depends on the chosen alignment, CIGAR, bases, clipping, mismatches, gaps, and scoring parameters. XS describes a competing alignment found during the reference search and contributes to placement confidence. MAPQ, NM, MD, SA, flags, and mate fields must describe the same alignment event. Sampling each field from real-read distributions could reproduce a histogram while creating impossible individual records. Starting with FASTQ lets one aligner calculate the fields jointly from the same sequence and reference.

Locked runtime

BWA source tag 0.7.12, source commit 5a340797e705d93ef0b0a0f0002eb198a9894e38; BWA binary SHA-256 a3fb8e9901543ac573e677316aa45c9fc7120c3ff0b88a7ed0c9e34822125309; hg19 FASTA SHA-256 92b96d16b307d824f3b6b9dc63feb237a75226f82ca2166d51ecec039bee4449.

5

Results

5.1 Primary recovery endpoint

All seven target fusions were detected in the spiked BAMs, and all seven matched backgrounds were negative within the locked target windows. Each spiked BAM contained exactly one matching candidate; no target window was ambiguous. The maximum absolute breakpoint error was 2 bp.

Table 2. Per-control results after whole-hg19 alignment. Error is left/right breakpoint offset. Split/span counts are unique FusionSieve supporting observations.
Control Fusion Pairs added Background target Spiked target Error (bp) Unique split/span Tier
Control 1EML4–ALK8001−2 / +251 / 20REVIEW
Control 2EML4–ALK8001−2 / +255 / 18REVIEW
Control 3EML4–ALK8001−2 / +256 / 20REVIEW
Control 4EML4–ALK8001−2 / +258 / 19REVIEW
Control 5EML4–ALK8001−2 / +259 / 20REVIEW
Control 6CD74–ROS17801+2 / +254 / 13REVIEW
Control 7CD74–ROS17501+2 / +252 / 11REVIEW
Total5530 / 7 controls7 / 7 controls≤ 2 absolute385 / 1210 PASS/HIGH

Concordance with the direct FASTQ mode

Relative to the earlier direct targeted-FASTQ analysis, target detection, reportability, breakpoint error, tier, score, and split/spanning support were identical in all seven controls. Whole-reference BAM ingestion enumerated fewer broad candidates: 31,565 to 1,024 across backgrounds and 31,572 to 1,031 across spiked datasets. The 97 background candidates sharing a target gene pair fell to zero. The two input modes enumerate background candidates differently, so the locked target endpoint and its associated call fields—not the broad candidate totals— are the directly concordant comparison. These changes are not an estimate of clinical specificity.

5.2 Alignment-field realism

The 553 planted pairs produced 1,106 primary and 417 secondary alignments. No planted primary record was unmapped. With the historical BWA -M behavior, shorter split hits were represented as secondary rather than supplementary alignments; this is expected aligner output.

Table 3. Aggregate audit of planted alignments across seven spiked BAMs.
MeasureObservedInterpretation
Mapped planted primary records1,106 / 1,106 (100%)Every planted read end received a primary placement
AS1,106 / 1,106 (100%)Calculated by BWA from each reported alignment
XS1,106 / 1,106 (100%)Calculated from competing placements in these records; not copied from real reads
NM, MD, and read group1,106 / 1,106 for each fieldAll required mapped-primary fields were present
SA417 / 1,106 (37.7%)Present where BWA emitted a linked chimeric/split representation; it is conditional, not universal
Secondary / supplementary records417 / 0Consistent with the selected BWA -M convention
Primary MAPQMedian 60; range 36–60Placement confidence was calculated by the aligner
Clipped primary records544 / 1,106 (49.2%)Clipping emerged from alignment rather than direct BAM construction
What changed from direct BAM insertion?

In the earlier direct BAM audit, 0% of planted records carried AS or XS. After FASTQ generation and BWA-MEM, 100% of mapped planted primary records carried both in this experiment. More importantly, the values now agree with the same alignment event that produced CIGAR, MAPQ, clipping, flags, mate fields, NM, MD, and conditional SA.

5.3 Comparison with length-matched background reads

All 1,106 planted primary records were compared with 1,106 ordinary primary records sampled deterministically from the matched background BAMs. Background records were stratified by control, read end, and 10-base query-length band. This controls basic read-length composition; it does not make an ordinary genomic read biologically equivalent to a fusion-junction read.

Table 4. Alignment structure in planted and length-matched background primary records.
MeasurePlanted fusion readsLength-matched background
Records1,1061,106
Median MAPQ6060
Any clipping49.2%2.4%
SA present37.7%0%
Proper-pair flag47.7%95.8%

The planted reads are intentionally chimeric. Their higher clipping and SA rates, and lower proper-pair rate, are expected manifestations of fusion structure. The comparison shows aligner-generated behavior; it does not claim that synthetic reads are identical to naturally occurring fusion molecules or ordinary background reads.

5.4 FusionSieve evidence

The seven target candidates contained 385 unique split reads and 121 unique spanning pairs, for 506 unique supporting observations. Each candidate had score 1.0 and no hard filters. Five candidates carried the additional soft filter low_depth; all seven carried reporting-readiness blockers described below. The score is a ranking score, not a probability of disease or a clinical confidence estimate.

5.5 Integrity checks

Table 5. Prepublication integrity summary.
CheckResult
Alignment manifests14 / 14 verified
FusionSieve rerun manifests and result JSON14 / 14 verified
Referenced file sizes and SHA-256 valuesAll verified
Matched backgrounds with zero target-window candidate7 / 7
Private identifiers exported in aggregate resultsNone

6

What REVIEW and PASS/HIGH mean

The phrase “none reached PASS or HIGH” does not mean the planted fusions were missed. All seven were detected and retained as REVIEW candidates. The terms describe two different decisions:

Table 6. Detection and automatic reportability are separate endpoints.
QuestionRule in this studyObserved result
Was the target detected?One REVIEW, PASS, or HIGH candidate within the locked window, no hard filters, and zero matched-background window callsYes: 7 / 7
Was it automatically reportable?The candidate must be PASS or HIGH and the reporting-readiness gate must be satisfiedNo: 0 / 7

All seven candidates were kept at REVIEW because automatic reporting was disabled, the heuristic scorer had not been independently validated, no compatible panel of normals was supplied, and duplicate-processing provenance could not be fully verified from reconstructed FASTQs. Controls 1–5 also had a soft low-depth flag at one side of the breakpoint. Controls 6–7 did not have that depth flag, but the same fail-closed readiness gate still applied.

Why the labels were not simply changed

Raising a REVIEW candidate to PASS/HIGH by editing a threshold or enabling a switch would not create the missing validation evidence. Honest automatic reporting requires a validated model, a compatible panel of normals, verified processing provenance, and a prospectively accepted reporting configuration.

7

Conclusions

  1. The missing full-reference alignment step was completed successfully for all 14 matched datasets.
  2. FASTQ-level generation followed by BWA-MEM removed the need to invent AS, XS, MAPQ, CIGAR, flags, or mate fields.
  3. FusionSieve recovered the intended target-window candidate in all seven spiked BAMs and in none of the matched backgrounds.
  4. The aligner-generated records passed the reported tag, sorting, index, and file-integrity checks.
  5. The candidates remain REVIEW-tier research findings. This study does not support automatic reporting or any clinical claim.

8

Limitations

  • The original instrument FASTQs were unavailable. FASTQ reconstruction cannot restore discarded duplicate molecules, original headers, raw file order, or pre-processing quality values.
  • Hard-clipped bases are absent from BAM records and cannot be reconstructed. This explains why some BAM-derived FASTQ query lengths are shorter than the original sequencer read length.
  • The deterministic reorder removes coordinate clustering but cannot recreate the historical flow-cell order or exact BWA batch boundaries.
  • This is a pinned BWA-MEM replay, not a byte-for-byte recreation of the unavailable historical Picard/GATK post-processing chain.
  • The hg19 FASTA is SHA-256 locked and contig-compatible with the source headers, but the historical BAM headers did not contain sequence MD5 fields that would prove whole-reference identity retrospectively.
  • Only seven backgrounds and two fusion definitions were evaluated. This is not a sensitivity, specificity, precision, reproducibility, or limit-of-detection study.
  • FusionSieve automatic reporting remained intentionally fail-closed because scorer, panel-of-normals, and processing-provenance requirements were not met.
  • No biological outcome, patient outcome, treatment response, or clinical decision was evaluated.

9

Future directions

  1. Recover original instrument FASTQs, if available, and repeat the replay without reconstruction.
  2. Reproduce the full historical post-processing chain when the exact Picard/GATK resources and intermediates can be verified.
  3. Expand the fusion panel across more genes, breakpoint architectures, read lengths, doses, and background depths.
  4. Run a blinded recovery study with truth locked before analysis and an independent caller.
  5. Validate reportability separately using a grouped training/validation design, a compatible panel of normals, and prospectively locked thresholds.
  6. Measure reproducibility across operating systems and production alignment environments.

10

Software and downloads

SpikeForge is the local research-control generator used to create the FASTQ-level fusion pairs. The public software site contains installation and use instructions. BAM and FASTQ files remain local because they are large and may contain sensitive sequencing information.

SpikeForge software

Download the research preview and validate it locally before use.

Open software and validation page

Full-reference follow-up package

Methods, aggregate results, alignment comparison, runtime and reference hashes, workflow scripts, tests, integrity manifest, and privacy statement. The 59,636-byte package excludes BAMs, FASTQs, indexes, read names, laboratory identifiers, local paths, runtime binaries, caller logs, and private manifests. Privacy scan: PASS with zero findings; archive CRC and inventory checks: PASS; packaged tests: 6 / 6 passed.

SHA-256: 4e0b3a377dacbff1c00070bfcb5f8ea1c9beef9ae935990e687f1dd4bd3d3d4f

Earlier FASTQ functional-study package

This existing package documents the preceding targeted-panel FASTQ study. It is retained so previously shared links continue to work. Use the full-reference package above for the current BWA-MEM replay and alignment-realism results.

SHA-256: 2161b446fcc7bbc1483ed37453a3fd4d9ba7625157edd3dad5b9c366313e9714

Related scientific reports

11

Technical references

  1. BWA reference implementation and documentation.
  2. SAM/BAM format specification.
  3. SAM optional fields specification.
  4. HTSlib and samtools documentation.