v2.2.1

salmon 2.2.1

The recommended 2.2.x release: it carries the full 2.2.0 feature set plus a
security fix for a lz4_flex advisory pulled in by the new RAD compression. If
you are on 2.2.0, upgrade. No index rebuild is required and quant.sf is
unchanged.

Security fix (the reason for the patch)

• lz4_flex advisory (GHSA-vvp9-7p8x-rfvv)
  — the RAD chunk de/compression introduced in 2.2.0 transitively pulled
  lz4_flex 0.10.0 (via libradicl), which carries a HIGH advisory: LZ4
  decompression could leak uninitialized memory or reuse the output buffer.
  salmon now depends on libradicl 0.14.1, which bumps lz4_flex to 0.13.1
  (the fix landed in 0.11.0). Drop-in: the RAD on-disk format is unchanged (LZ4's
  block format is stable, so RAD files written by 2.2.0 still decompress) and the
  compressed-RAD round-trip stays byte-identical.

---

The rest of this note summarizes the 2.2 feature set (introduced in 2.2.0 and
shipped here), since 2.2.1 is the release most users will land on.

Decoupled mapping and quantification (--writeRad, --rad)

salmon can now separate mapping from quantification through the RAD format:

• --writeRad <PATH> writes per-fragment mappings (sketch or
  selective-alignment profile, chosen automatically). Quantification still runs;
  add --skipQuant to map only. The file is piscem map-bulk-compatible.
• --rad <PATH> quantifies a RAD file directly and in parallel — salmon-
  written or piscem map-bulk/sketch output. No -i is needed; reference
  names travel in the RAD header.

This lets you map once and quantify many times, quantify a shared piscem RAD with
salmon's EM, or split the two phases across machines. Quantifying a
piscem-produced RAD is fast and lean — 16.6 s and 0.52 GB RSS for the quant
half of a 36 M-read human sample (mapping already done by piscem).

Deterministic quantification (--deterministic) — reproducible and faster

RAD-mode quantification is byte-identical across thread counts and runs, and
--deterministic brings that guarantee to FASTQ input: it maps the reads once to
an intermediate RAD, then quantifies from it with a fixed fragment-length
distribution. Determinism comes from making the computation order-independent
(fixed-point u128 equivalence-class weight accumulation; integer-count
fragment-length histograms built once in a fixed order) rather than from sorting
records.

It is reproducibility at negative cost. Because the single mapping pass
feeds a fast deterministic streaming requant instead of the heavier online
dual-phase inference, --deterministic is consistently faster and lighter
than the default online path. Measured on a 36 M-read GEUVADIS sample
(ERR188044, paired-end) against a decoy-aware human index, 16 threads:

run	wall	peak RSS
C++ salmon 1.9 / 2.1.2 (selective alignment)	97.9 s	9.9 GB
salmon 2.2 direct (selective alignment)	94.7 s	10.0 GB
salmon 2.2 --deterministic (selective alignment)	90.7 s	9.8 GB
salmon 2.2 direct (sketch)	53.1 s	9.8 GB
salmon 2.2 --deterministic (sketch)	45.2 s	9.7 GB

So --deterministic runs ~4% faster than the default in selective-alignment
mode and ~15% faster in sketch mode (and ~7% faster than C++ salmon in SA),
while also using slightly less memory — and unlike either non-deterministic path,
its quant.sf is bit-for-bit identical regardless of -p. Reproducibility is no
longer a speed/accuracy tradeoff.

--deterministic works with bias correction and does not require -t (the
reference sequences for the second pass are taken from the index). The
intermediate RAD is removed on success unless --keepRad (or --writeRad PATH).

Bias-aware RAD, and deterministic + faster bias models

Sequence, GC, and positional bias correction now work on RAD input
(--rad … --seqBias/--gcBias/--posBias), sharing the same correction tail as the
alignment (BAM) path.

• Deterministic bias models. Bias observed models accumulate per-fragment
  mass in fixed-point integers, so the trained models — and the bias-corrected
  effective lengths — are byte-identical across thread counts.
• Faster bias correction. The sequence-bias per-position factor build folds
  the observed/expected models into a single precomputed obs − exp log-bias
  table (one context encode + one table sweep instead of two), giving ~1.9× on
  the seqBias correction sweep; the length-sweep convolution for any non-GC
  bias is computed once as an FFT cross-correlation (O(L log L)). Both are
  reassociation-level and accuracy-vs-truth is unchanged.

RAD chunk compression (--radCompress)

RAD output (and the --deterministic intermediate) can be compressed per chunk:
--radCompress=lz4|zstd|none (default lz4), --noCompressRad to force
uncompressed. Transparent and lossless — chunks are decompressed in the reader,
so all consumers (salmon, alevin-fry, piscem-infer) are unchanged, and a RAD with
no codec tag (any pre-2.2 / piscem file) reads as uncompressed automatically. On a
36 M-read human RAD, lz4 is ≈ 1.25× smaller and zstd ≈ 1.9×, with the
requantified quant.sf byte-identical regardless of codec.

Compatibility

• No index rebuild. On-disk index format is unchanged from 2.1.x; quant.sf
  and the inferential-replicate formats are unchanged.
• A plain salmon quant -i … -1 … -2 … run is unaffected, except that the bias
  models are now thread-count-deterministic (accuracy vs truth unchanged).

Attribution

The RAD direction was opened by @BenjaminDEMAILLE in PRs
#1033,
#1034, and
#1037, which the 2.2 series
supersedes. It adopts his core design — decouple map/quant through RAD, a
--writeRad producer and --rad consumer, and piscem bulk/sketch RAD as a
first-class input — and differs in the determinism mechanism (order-independent
computation + a fixed, baked FLD rather than an external merge sort) plus the
information-adaptive 1-vs-2-pass dispatch, bias-aware RAD, and chunk compression.

---

Install salmon-cli 2.2.1

Install prebuilt binaries via shell script

curl --proto '=https' --tlsv1.2 -LsSf https://github.com/COMBINE-lab/salmon/releases/download/v2.2.1/salmon-cli-installer.sh | sh

Download salmon-cli 2.2.1

File	Platform	Checksum
salmon-cli-aarch64-apple-darwin.tar.xz	Apple Silicon macOS	checksum
salmon-cli-x86_64-apple-darwin.tar.xz	Intel macOS	checksum
salmon-cli-aarch64-unknown-linux-gnu.tar.xz	ARM64 Linux	checksum
salmon-cli-x86_64-unknown-linux-gnu.tar.xz	x64 Linux	checksum