apex: clean apex-profile weighting (bench-validated, perf-optimized) by jspaezp · Pull Request #68 · TalusBio/timsbuktoolkit

jspaezp · 2026-07-19T00:52:52Z

Clean re-implementation of the apex-profile weighting experiments from the
apex-autoresearch branch, rebuilt off main as maintainable, tested, and
runtime-optimized code. Every term was bench-validated on the apex_sim
canonical suite; only terms that clear the agreed keep-bar ship.

What changed

Base apex profile retuned and two per-cycle weighting passes added, applied to
apex_profile before peak-picking:

Base: cos^0.5 * I^0.75 * (1.75 + s_norm), one gaussian-blur pass, joint-apex
snap window 1 (was cos^3 * I, floor 0.5, no blur, snap 3).
Coelution weight: ratio-free expected-intensity-weighted mean pairwise
correlation of center-normalized fragment XIC windows. Rewards genuine
co-elution, independent of absolute fragment ratios.
Vote weight: robust per-row matched-filter soft votes summed across
fragments; saturating votes let a real multi-fragment apex win over tall
single-row interferents.

All knobs live in one ApexConfig (bench-tuned Default, no env-var globals). The
passes are pure kernels in scoring/apex/{coelution,vote}.rs over
(profile, rows, cfg, scratch), decoupled from the extraction data model and
unit-tested.

Ablation (keep-bar: SUM gain > 2.0 on the 6-scenario suite; noise = 0)

term decision SUM
base retune KEEP 267.9 -> 326.2 (+58.3)
gaussblur (1 pass) KEEP included above
joint-snap 1 KEEP +1.5 (free: existing constant, no new code)
coelution weight KEEP 326.2 -> 341.5 (+15.3)
MS1 isotope-coherence DROP +1.3 (< 2.0; regressed mismatched, and
dropping it keeps the stack
precursor-independent)
vote weight KEEP 341.5 -> 405.1 (+63.6)

Rounded exponents/strengths (0.5 / 0.75 / 1.75; coel_k 1.0, vote_k 14, tau 1.3,
s 0.4) match or beat the raw Optuna values, so no over-precise magic constants
ship. Also added an absent_precursor bench scenario as a permanent guard
against over-reliance on precursor signal.

Runtime

The passes run per-candidate on rayon workers. Cost recovered from 91 to 51
us/run, result-identical (SUM unchanged at each step):

vote median/MAD: 3 full sorts/row -> select_nth (O(n)). 91 -> 59
coelution: pairwise O(rows^2) -> ||W||^2 closed form (O(rows))
- skip zero-profile cycles. 59 -> 51
per-worker scratch: no per-candidate allocation. parity

vs apex-autoresearch branch

branch (isotope on, unoptimized): SUM 494.7 @ 91.8 us/run
this PR (isotope dropped, opt): SUM 491.6 @ 51.2 us/run

-0.7% SUM for +44% throughput, a simpler precursor-independent stack, 3 fewer
knobs, and unit tests. The 3.4-point delta is entirely the dropped isotope term.

Validation

cargo test -p timsseek: 94 passed (4 new apex unit tests).
cargo build -p timsseek --release, cargo build -p apex_sim --release: clean.
No env/LazyLock knobs remain (grep clean).
Bench: bash bench_out/score.sh (apex_sim canonical suite, synthetic).

…blur pass, joint-snap 1) Bench-validated: SUM 267.9 -> 326.2 (+58.3) on apex_sim canonical suite. Rounded exponents (0.5/0.75/1.75) match or beat the Optuna values.

Per-cycle expected-intensity-weighted mean pairwise correlation of center-normalized fragment XIC windows. Runtime cost (18->44 us/run) addressed in a later perf pass.

…, +63.6) Robust per-row matched-filter soft votes summed across fragments; saturating votes make a genuine multi-fragment apex win over tall single-row interferents. Runtime cost (44->91 us/run) addressed in perf pass.

MS1 precursor undetected, fragments intact. Guards against over-reliance on precursor signal in the apex profile.

Replaces 3 full sorts/row with partial selection. Result-identical (same order statistic); SUM 491.3 unchanged on canonical suite.

… 59->51 us/run) Weighted mean pairwise correlation of unit window vectors has a closed form (sum_{a<b} w_a w_b <u_a,u_b> = (||W||^2 - sum w_a^2)/2), collapsing the per-cycle O(rows^2) inner loop to O(rows). Zero-profile cycles skip the window work entirely. Result-identical: SUM 491.3 unchanged.

…date alloc) Moves cn/active/wacc and support/r/m/tmp into an ApexScratch owned by TraceScorer, cleared+resized per call. Matches the TraceScorerBuffers idiom; removes allocator churn across rayon workers at scale. Result-identical: SUM 491.3 unchanged.

Moves coelution + vote into scoring/apex/{coelution,vote}.rs as pure kernels over (profile, rows, cfg, scratch), decoupled from the Extraction data model. ApexConfig/ApexScratch live in apex/mod.rs; apex_finding calls a single weight_profile orchestrator. Adds 4 unit tests (co-elution vs scattered, concurrent votes vs tall interferent, disabled no-ops). Bench parity: SUM 491.3, 51 us/run unchanged.

…, s 0.4) Rounded from raw Optuna values; SUM 491.3 -> 491.6, no regression.

github-actions · 2026-07-19T00:57:10Z

Apex-finder bench

cargo run -p apex_sim --release --example bench -- 1000 2

Sensitivity + timing across canonical scenarios

=== clean ===
  config: n_cycles=245 sigma=1 noise_floor=0.02 random_peaks=200 true_apex=30
  runs=1000 tol=±2 cycles
  build :    21.81 ms total  (21.81 us/run)
  score :    69.53 ms total  (69.53 us/run, 14382 runs/s)
  scorer errors:    0
  pass1 sensitivity: 100.00%  (1000/1000)
  pass2 sensitivity: 100.00%  (1000/1000)
  pass2 |apex-truth| cycles: mean=0.00 median=0 max=0
  pass2 err histogram: 0:1000 1:0 2:0 3:0 4:0 5:0 6:0 7+:0 

=== moderate_noise ===
  config: n_cycles=245 sigma=1 noise_floor=0.2 random_peaks=50 true_apex=30
  runs=1000 tol=±2 cycles
  build :    71.15 ms total  (71.15 us/run)
  score :    69.99 ms total  (69.99 us/run, 14287 runs/s)
  scorer errors:    0
  pass1 sensitivity:  99.90%  (999/1000)
  pass2 sensitivity:  99.80%  (998/1000)
  pass2 |apex-truth| cycles: mean=0.64 median=1 max=41
  pass2 err histogram: 0:451 1:503 2:44 3:1 4:0 5:0 6:0 7+:1 

=== high_noise+interference ===
  config: n_cycles=245 sigma=1 noise_floor=0.5 random_peaks=100 true_apex=30
  runs=1000 tol=±2 cycles
  build :   124.98 ms total  (124.98 us/run)
  score :    70.92 ms total  (70.92 us/run, 14100 runs/s)
  scorer errors:    0
  pass1 sensitivity:  86.50%  (865/1000)
  pass2 sensitivity:  86.40%  (864/1000)
  pass2 |apex-truth| cycles: mean=13.81 median=1 max=213
  pass2 err histogram: 0:316 1:495 2:53 3:4 4:1 5:1 6:3 7+:127 

=== heavy_interference ===
  config: n_cycles=245 sigma=1 noise_floor=0.5 random_peaks=300 true_apex=30
  runs=1000 tol=±2 cycles
  build :   346.15 ms total  (346.15 us/run)
  score :    70.84 ms total  (70.84 us/run, 14116 runs/s)
  scorer errors:    0
  pass1 sensitivity:   4.00%  (40/1000)
  pass2 sensitivity:   4.00%  (40/1000)
  pass2 |apex-truth| cycles: mean=92.09 median=87 max=214
  pass2 err histogram: 0:15 1:14 2:11 3:6 4:8 5:12 6:8 7+:926 

=== mismatched_library ===
  config: n_cycles=245 sigma=1 noise_floor=0.5 random_peaks=100 true_apex=30
  runs=1000 tol=±2 cycles
  build :   125.42 ms total  (125.42 us/run)
  score :    71.07 ms total  (71.07 us/run, 14072 runs/s)
  scorer errors:    0
  pass1 sensitivity:  88.80%  (888/1000)
  pass2 sensitivity:  88.80%  (888/1000)
  pass2 |apex-truth| cycles: mean=11.29 median=1 max=211
  pass2 err histogram: 0:317 1:519 2:52 3:1 4:0 5:1 6:2 7+:108 

=== absent_top_fragment ===
  config: n_cycles=245 sigma=1 noise_floor=0.5 random_peaks=100 true_apex=30
  runs=1000 tol=±2 cycles
  build :   123.33 ms total  (123.33 us/run)
  score :    69.65 ms total  (69.65 us/run, 14357 runs/s)
  scorer errors:    0
  pass1 sensitivity:  26.30%  (263/1000)
  pass2 sensitivity:  26.30%  (263/1000)
  pass2 |apex-truth| cycles: mean=70.83 median=49 max=213
  pass2 err histogram: 0:108 1:124 2:31 3:4 4:5 5:5 6:3 7+:720 

=== absent_precursor ===
  config: n_cycles=245 sigma=1 noise_floor=0.5 random_peaks=100 true_apex=30
  runs=1000 tol=±2 cycles
  build :   122.92 ms total  (122.92 us/run)
  score :    69.37 ms total  (69.37 us/run, 14414 runs/s)
  scorer errors:    0
  pass1 sensitivity:  86.50%  (865/1000)
  pass2 sensitivity:  86.40%  (864/1000)
  pass2 |apex-truth| cycles: mean=13.82 median=1 max=213
  pass2 err histogram: 0:306 1:504 2:54 3:4 4:1 5:1 6:3 7+:127 

=== summary (n=1000, tol=±2 cycles) ===
  scenario                    pass2%  pass1%   medErr    us/run
  clean                        100.0   100.0        0     69.53
  moderate_noise                99.8    99.9        1     69.99
  high_noise+interference       86.4    86.5        1     70.92
  heavy_interference             4.0     4.0       87     70.84
  mismatched_library            88.8    88.8        1     71.07
  absent_top_fragment           26.3    26.3       49     69.65
  absent_precursor              86.4    86.5        1     69.37

commit 3130a97

…_loop) Result-identical; SUM 491.6 unchanged.

jspaezp added 10 commits July 18, 2026 17:26

apex: add ApexConfig scaffold (default reproduces main)

648ad3b

apex: retune base apex profile (cos^0.5*I^0.75, scribe floor 1.75, 1 …

c6b9440

…blur pass, joint-snap 1) Bench-validated: SUM 267.9 -> 326.2 (+58.3) on apex_sim canonical suite. Rounded exponents (0.5/0.75/1.75) match or beat the Optuna values.

apex: add ratio-free coelution weight (KEEP: SUM 326.2 -> 341.5, +15.3)

27bd8b1

Per-cycle expected-intensity-weighted mean pairwise correlation of center-normalized fragment XIC windows. Runtime cost (18->44 us/run) addressed in a later perf pass.

apex_sim(bench): add absent_precursor scenario

37ef77f

MS1 precursor undetected, fragments intact. Guards against over-reliance on precursor signal in the apex profile.

apex: O(n) median/MAD via select_nth in vote pass (perf: 91->59 us/run)

0d7215c

Replaces 3 full sorts/row with partial selection. Result-identical (same order statistic); SUM 491.3 unchanged on canonical suite.

apex: round weight-strength constants (coel_k 1.0, vote_k 14, tau 1.3…

075657d

…, s 0.4) Rounded from raw Optuna values; SUM 491.3 -> 491.6, no regression.

apex: iterate profile/m by ref in weight loops (clippy needless_range…

db8209b

…_loop) Result-identical; SUM 491.6 unchanged.

This was referenced Jul 19, 2026

apex_sim: production-realistic bench conditions + score-discrimination (AUC) #69

Open

apex: apex_sim bench + unify finder + cull weighting, re-tune base profile #70

Merged

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

apex: clean apex-profile weighting (bench-validated, perf-optimized)#68

apex: clean apex-profile weighting (bench-validated, perf-optimized)#68
jspaezp wants to merge 11 commits into
mainfrom
apex-weighting-clean

jspaezp commented Jul 19, 2026

Uh oh!

github-actions Bot commented Jul 19, 2026 •

edited

Loading

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

Uh oh!

Conversation

jspaezp commented Jul 19, 2026

What changed

Ablation (keep-bar: SUM gain > 2.0 on the 6-scenario suite; noise = 0)

Runtime

vs apex-autoresearch branch

Validation

Uh oh!

github-actions Bot commented Jul 19, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Apex-finder bench

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

1 participant

github-actions Bot commented Jul 19, 2026 •

edited

Loading