STAR compatibility

rustar-aligner is a faithful port. The goal is byte-for-byte identical output to STAR for every read where the algorithm is deterministic, and provably equivalent output (same alignment set, different tie-break) for every read where it isn’t. This page is the long-form scoreboard.

The benchmark below uses 10,000 yeast RNA-seq reads (150 bp, ERR12389696), aligned with both tools using identical parameters and the same genome index.

Single-end summary

Metric	rustar-aligner	STAR
Unique mapped	82.6%	82.6%
Multi-mapped	7.4%	7.4%
Total mapped	90.0%	90.0%
Position agreement (raw)	96.5%	—
Position agreement (tie-adjusted)	99.815%	—
Reads mapped only by STAR	0	—
Reads mapped only by rustar-aligner	0	—
CIGAR-only differences	1	—

Tie-adjusted means: of the 313 raw disagreements, 299 are verified genuine ties — both tools find the same set of alignments, but pick different copies as primary because of differences in suffix-array iteration order or RNG-based tie-breaking. Excluding those ties, faithfulness is 8,611 / 8,627 = 99.815%.

Paired-end summary

Metric	rustar-aligner	STAR
Both mates mapped	8,390	8,390
Half-mapped pairs	0	0
Unmapped pairs	0	0
PE faithfulness (tie-adjusted)	99.883%	—
MAPQ inflations	0	—
MAPQ deflations	0	—
NH tag differences	0	—
Proper-pair flag differences	0	—

Tie-adjusted for paired-end: 16,284 / 16,306 mate alignments exactly match STAR (same position, CIGAR, MAPQ, proper-pair flag, NH tag). 475 differences are excluded as tie-breaking only (same MAPQ + same NH, different repeat copy chosen).

Index format

The genome index format is identical. After Phase G3 (the SA tie-breaking fix), the suffix array for the yeast benchmark genome is byte-for-byte identical between STAR and rustar-aligner: 10,862 entry differences fixed → 0 remaining.

This means an index built with one tool is loadable by the other.

Where the differences come from

There are three categories of remaining difference, in order of size:

1. Tie-breaking (the bulk: ~3% of reads)

When two alignments have the same score, both tools pick a primary using a deterministic but different procedure:

• STAR uses a Mersenne Twister RNG (mt19937) seeded by --runRNGseed.
• rustar-aligner uses Rust’s StdRng (ChaCha) seeded by --runRNGseed.

Both honour the seed and are reproducible across runs of the same tool — but the produced sequences differ between tools. So for ~3% of multi-mapped reads, primary vs. secondary flips. Total NH counts, AS scores, CIGARs, and the full alignment set are unaffected.

A subset (~100 of 313 SE diffs) involves “diff-chr ties”: the read maps to several copies in a multi-copy region (e.g. rDNA), and the two tools pick different copies. Same alignment quality, different one chosen.

This is what the “tie-adjusted” numbers above account for.

2. CIGAR placement in homopolymer runs (1 read in 10,000)

ERR12389696.13573895: both tools align to XV:218357 MAPQ=255 AS=133, but rustar-aligner emits 100M1I45M4S (insertion at read position 100) while STAR emits 108M1I37M4S (insertion at 108). The 71-base seed is found at RC pos 29 (rustar-aligner) vs RC pos 37 (STAR) due to a different Lmapped chain path through a long homopolymer. Same diagonal, same score, different starting position → different insertion placement.

This is a seed-level tie. Real impact for downstream tools: effectively zero. To match STAR exactly we’d need to replicate STAR’s exact Lmapped chain, which is a high-effort, low-value fix.

3. PE-specific cases where rustar-aligner finds a better alignment

Four PE alignments have a higher AS in rustar-aligner than STAR. These are not bugs — they’re improvements:

• ERR12389696.844151: rustar-aligner finds VIII:451791 with 0 mismatches; STAR finds VII:1001391 with 6 mismatches.
• ERR12389696.4972950: rustar-aligner correctly emits a spliced mate 2; STAR’s combined-window approach fails to stitch the spliced mate at the better location and emits unspliced.

In both, STAR’s combined-window seeding fails to reach the higher-scoring alignment. We’ve decided not to artificially regress these.

Faithfulness over time

The faithfulness numbers have been moving upwards as more STAR algorithm details have been replicated:

Phase	PE tie-adjusted faithfulness
Pre-Phase F1	99.755%
Phase F1 (--runRNGseed)	99.755%* (RNG change reset baseline)
Phase G1 (junction-shift fix)	99.865%
Phase G2 (recursion budget + overflow fix)	99.883%

* Phase F1 changed PE tie-breaking from SA-order to seeded StdRng, which shuffled which reads count as “tie-broken” without changing the underlying alignment quality.

Other compatibility checks

• Genome index files (Genome, SA, SAindex, chrName.txt, chrStart.txt, chrNameLength.txt, sjdbList.fromGTF.out.tab, transcriptInfo.tab, exonInfo.tab): identical or equivalent.
• SJ.out.tab: matching format and contents for the yeast benchmark.
• Log.final.out: format matches STAR; MultiQC parses it without modification.
• Chimeric.out.junction: 14-column STAR-compatible format. Tools like Arriba and STAR-Fusion read it directly.
• ReadsPerGene.out.tab: 4-column STAR-compatible format.

What this means in practice

For the overwhelming majority of bioinformatics workflows — read counting, differential expression, splice junction analysis, fusion calling, MultiQC reports — rustar-aligner’s output is interchangeable with STAR’s. The exceptions are:

• Workflows that depend on the specific copy a multi-mapper landed on (rare; in those cases, you should be using the full alignment set anyway, not just the primary).
• Reproducing exact byte-equality of SAM output across STAR and rustar-aligner runs (not achievable today; the RNG difference is the dominant source).

If you find a divergence not described here, please open an issue — the project goal is to keep this list as short as possible.