Gaolaboratory/chimera2

Chimera study 2: fine-tuning (post-training) and test data

Deposited 2026-11-12 from /mnt/powervault/chimeric_decompose/experiments/chimera2_20261112/. Built by scripts/c2_build_dda.py and scripts/c2_build_dia.py in that repo; read by scripts/c2_train.py and scripts/c2_eval.py. Design: docs/superpowers/specs/2026-11-12-chimera2-design.md.

All files are numpy .npz. Spectra are stored ragged: spectrum i owns peaks offsets[i]:offsets[i+1] of mz (float64, sorted) and inten (float32); the most intense peaks of each MS2 scan are kept. Per-spectrum arrays: iso_lo, iso_hi (isolation window, Th), sel_mz, charge, scan, rt_min, n_peaks and tic (of the original scan), run (index into info.json runs).

dda_chimericity/ (task A: is a DDA MS2 spectrum chimeric?)

Source: public PRIDE human label-free Q Exactive / Exploris DDA projects, 3 runs per project (selected_runs.json). Labels from a pfsearch DDA+ search (lib_richmod_v1, rescored), per-run target-decoy q. label 1 = at least 2 distinct peptide backbones at q <= 0.01 (mods stripped, I = L); 0 = exactly one. Spectra with an accepted entrapment hit are dropped. Split by project (hash): about 70 / 10 / 20.

file	spectra	notes
train.npz	3,097,657	post-training set; all positives (1,255,824) plus 3 negatives per positive per run
val.npz	437,124	natural prevalence, up to 15,000 per run
test.npz	702,356	natural prevalence (221,054 positive), up to 15,000 per run
test_unid.npz	338,073	test-project spectra with no identification at 1% (secondary analysis; no label)
eval_subset/val.npz	58,920	fixed-seed subsample, up to 1,500 per run; used for model selection
eval_subset/test.npz	182,466	fixed-seed subsample, up to 3,000 per run; used for reported test numbers
info.json		run list (project, run, split) and per-run label statistics

dia_precursor/ (task B: which precursors are in a DIA MS2 spectrum?)

Source: PXD003179 HeLa, 5 / 10 / 15 / 20 / 25 Da windows, 3 replicates each (15 runs). Labels from rs-diann in library mode (DIA-NN speclib diann_entrap_q05/report-lib.predicted.speclib): a precursor belongs to a scan when its m/z is inside the isolation window and the scan RT is within RT.Start..RT.Stop. Replicate 03 of every width is test; replicates 01 and 02 are train; 4% of replicate 02 is validation.

Labels are ragged: spectrum i owns labels lab_offsets[i]:lab_offsets[i+1]. lab_mz, lab_charge: the precursor. lab_q: target index = charge_index * 52 + bin, charge in (2, 3), bin = 0.5 Th steps from iso_lo. lab_kind: 1 = human precursor at Q <= 0.01 (positive); 2 = 0.01 < Q <= 0.05 or non-human (ignored in loss and scoring). lab_held: peptide belongs to the 20% held out by sequence hash (masked from the training loss; scored separately at test). width: nominal window width in Da.

file	spectra	labels
train.npz	1,180,827	5,425,425
val.npz	24,000	112,256
test.npz	601,719	2,740,251
eval_subset/test.npz	100,000	fixed-seed subsample, 20,000 per test run; used for reported test numbers
info.json		per-run precursor counts and split