Merge pull request #1023 from COMBINE-lab/feat/no-frag-length-dist

feat(quant): implement --noFragLengthDist; confirm sketch fragment-length weighting

Author: rob-p

crates/salmon-cli/src/main.rs

@@ -708,9 +708,6 @@ fn run_quant(args: QuantArgs, quiet: bool) -> Result<()> {
     if args.eqclasses.is_some() {
         tracing::warn!("--eqclasses (quantify from a precomputed equivalence-class file) is not yet implemented and is ignored; mapping/alignment input is used instead.");
     }
-    if args.no_frag_length_dist {
-        tracing::warn!("--noFragLengthDist is accepted but not yet implemented and has no effect: the fragment-length distribution is still used in the per-fragment probability.");
-    }
     if args.sample_out || args.sample_unaligned || args.write_qualities {
         tracing::warn!("--sampleOut/--sampleUnaligned/--writeQualities (posterior-sampled BAM output) are accepted but not yet implemented and have no effect.");
     }
@@ -866,6 +863,7 @@ fn run_quant(args: QuantArgs, quiet: bool) -> Result<()> {
     opts.thinning_factor = args.thinning_factor;
     opts.no_length_correction = args.no_length_correction;
     opts.model_single_frag_prob = !args.no_single_frag_prob;
+    opts.no_frag_length_dist = args.no_frag_length_dist;
     opts.map_config.align.min_score_fraction = args.min_score_fraction;
     opts.map_config.pair.orphan_chain_sub_thresh = args.orphan_chain_sub_thresh;
     opts.map_config.align.full_length_alignment = args.full_length_alignment;

crates/salmon-quant/src/lib.rs

@@ -101,6 +101,9 @@ pub struct QuantOptions {
     /// the bounded-CMF "ambiguous" weight (salmon default `true`); `false` =
     /// `--noSingleFragProb`.
     pub model_single_frag_prob: bool,
+    /// disable the fragment-length distribution in the per-fragment assignment
+    /// probability (`--noFragLengthDist`); default `false`.
+    pub no_frag_length_dist: bool,
     /// fragment-length distribution prior mean, SD, and max tracked length
     /// (`--fldMean` / `--fldSD` / `--fldMax`)
     pub fld_mean: f64,
@@ -173,6 +176,7 @@ impl QuantOptions {
             thinning_factor: 16,
             no_length_correction: false,
             model_single_frag_prob: true,
+            no_frag_length_dist: false,
             fld_mean: 250.0,
             fld_sd: 25.0,
             fld_max: 1000,
@@ -422,6 +426,7 @@ pub fn quantify(opts: &QuantOptions) -> Result<QuantResult> {
             online: online.as_ref(),
             paired_lib: opts.is_paired(),
             model_single_frag_prob: opts.model_single_frag_prob,
+            no_frag_length_dist: opts.no_frag_length_dist,
             num_processed,
             num_mapped,
             num_orphan: &num_orphan,

crates/salmon-quant/src/processor.rs

@@ -87,6 +87,9 @@ pub(crate) struct Shared<'a> {
     /// the bounded-CMF "ambiguous" weight (salmon default). When `false`
     /// (`--noSingleFragProb`) orphans fall back to a flat penalty / weight-1.
     pub model_single_frag_prob: bool,
+    /// disable the fragment-length distribution in the per-fragment assignment
+    /// probability entirely (`--noFragLengthDist`).
+    pub no_frag_length_dist: bool,
     pub num_processed: &'a AtomicU64,
     pub num_mapped: &'a AtomicU64,
     /// mapped fragments whose representative mapping is an orphan (only one mate
@@ -196,9 +199,16 @@ fn frag_log_prob(
     use_aux: bool,
     model_single_frag_prob: bool,
     paired_lib: bool,
+    no_frag_length_dist: bool,
     pmf: &[f64],
     cmf: &[f64],
 ) -> f64 {
+    // --noFragLengthDist: do not consider the fragment-length distribution in the
+    // per-fragment assignment probability (salmon's flag). Applies to both the
+    // proper-pair PMF term and the orphan/single-end ambiguous term.
+    if no_frag_length_dist {
+        return LOG_1;
+    }
     if m.status == MateStatus::PairedEndPaired && m.fragment_len > 0 {
         if use_aux && !pmf.is_empty() {
             let flen = (m.fragment_len as usize).min(pmf.len() - 1);
@@ -461,6 +471,7 @@ fn record(
                     use_aux,
                     sh.model_single_frag_prob,
                     sh.paired_lib,
+                    sh.no_frag_length_dist,
                     &fld_pmf,
                     &fld_cmf,
                 );
@@ -529,6 +540,7 @@ fn record(
                 use_aux,
                 sh.model_single_frag_prob,
                 sh.paired_lib,
+                sh.no_frag_length_dist,
                 &fld_pmf,
                 &fld_cmf,
             )

docs/release-notes-2.1.0.md

@@ -356,6 +356,17 @@ Proper pairs are likewise length-conditioned (`pmf(flen) − cmf(txpLen)`), whic
 transcripts. The new `--noSingleFragProb` flag (default off) restores the old flat
 behavior, matching salmon's option of the same name.
 
+This fragment-length weighting now applies to **sketch (pseudoalignment) mode**
+too: sketch already trains the fragment-length distribution from concordant
+pairs, and — now that sketch mappings carry their positions (above) — the
+orphan/single-end ambiguous term contributes as well. `--noFragLengthDist`,
+previously an accept-and-warn no-op, is **now implemented**: it disables the
+fragment-length term in the per-fragment assignment probability for both modes.
+On the simulated data, enabling the fragment-length model (the default) improves
+accuracy in both modes — Spearman vs truth rises ~0.011 (SA) / ~0.012 (sketch) on
+the easy set and ~0.015 / ~0.016 on the hard set — and brings sketch accuracy to
+within ~0.005 Spearman of selective alignment.
+
 ## `num_dovetail_fragments` is now reported
 
 Rust always dropped dovetailed concordant pairs under the default no-dovetail