CID fragmentation physics and collision-energy semantics for timsTOF vs Orbitrap vs QTOF (CASMI 2026)

TL;DR. The hidden test is a Bruker timsTOF at fixed lab-frame collision energies of 20, 40 and 60 eV; one quarter of the spectra are "merged", and the enveda-180 code plus a 600-group numerical test show a merged spectrum is the raw-ion-count sum of the three single-energy scans (re-centroided), so it is dominated by the lower-energy scans and looks like a mid-energy (Orbitrap NCE 30-45) spectrum. The train parquet converts Orbitrap NCE to eV with eV = NCE x (m/z)/500 (charge factor 1), passes through QTOF volts as eV, and leaves ion-trap NCE unconverted; but empirically a timsTOF 20 eV scan matches an Orbitrap NCE 20 scan (median sqrt-cosine 0.82), not the 33 NCE that the formula implies, so every NCE-conditioned model or Orbitrap library must be driven with ~0.65-0.75x the timsTOF lab eV. The forced astrophysics analogy (forward model x instrument response) yields one concrete, fully specified experiment with 2,421 timsTOF/Orbitrap training pairs and a scaffold-split MRR@25 evaluation; it partially holds.

1. How collision energy is set on each platform

1.1 Bruker timsTOF (PASEF)

In dda-PASEF the quadrupole isolates each precursor as it elutes from the TIMS ramp, and the method file maps collision energy to the ion-mobility coordinate 1/K0, not directly to m/z (m/z and 1/K0 are correlated, so this is an indirect mass scaling). Verified settings:

I could not open the enveda-180 preprint (README cites ChemRxiv DOI 10.26434/chemrxiv.15004319/v1 - the DOI pattern looks non-standard and did not resolve for me; unverified), so isolation width, whether the three energies were separate injections or in-run CE stepping, and the TIMS ramp are open. Bruker's small-molecule default is thus not one number: the Metabolomics "collision energy stepping" idea exists (STAR protocol above) but Enveda chose fixed energies.

1.2 Lab-frame vs centre-of-mass energy

Only the centre-of-mass fraction of the lab energy can be converted to internal energy in one collision: E_com = E_lab x m_gas/(m_gas + m_ion). Bruker QTOF/timsTOF collision cells and Thermo HCD cells both use N2 (Szabo et al. 2021: "The collision gas in both instruments is nitrogen"); Waters QTOFs typically use argon.

Single-collision E_com is far below bond energies (3-4 eV); dissociation needs multiple collisions (several in a 10^-2 mbar cell), so the mapping from lab eV to internal energy is instrument-specific (gas pressure, cell length, residence time). This is why Szabo et al. find a slope of 0.92 (R2 = 0.997) between best-matching QTof-CID and Orbitrap-HCD energies for leucine enkephalin ("1 eV increase in collision energy on the Thermo Orbitrap increases the internal energy of the precursor ion slightly more than 1 eV increase on the Bruker QTof"), with HCD needing "a few eV lower" (QTof 20/30/60 eV matched by HCD 18/28/55 eV; peptide example QTof 50 eV ~ HCD 42 eV).

1.3 Thermo normalized collision energy (NCE)

Thermo's PSB104 bulletin defines NCE as a 0-100% scale that "automatically compensates for this mass dependency" because "the collision energy needed to achieve optimum fragmentation efficiency has been shown to follow a linear correlation with m/z". The de facto formula for HCD is eV = NCE x (m/z)/500 x f(z); the charge factors f(2) = 0.9, f(3) = 0.85 appear in Szabo et al. 2021 (their Bruker default-CE equation, "The charge factor equals to 0.9 and 0.85 in the case of doubly and triply charged peptides"); a patent-text variant CE = NCE/100 x (0.2 - z/100) x m/z surfaced in search snippets only (low confidence, not read in the original). What matters for us: the train parquet used exactly eV = NCE x (m/z)/500 - the ratio ev/(NCE x m/z/500) over 966,330 pluskal_ms2 energy values has median 0.9998 (e.g. NCE 15/20/60 at m/z 468.1444 -> 14.0/18.7/56.2 eV).

2. Semantics of collision_energy_ev, collision_energy_orig, collision_energy_orig_units (computed from train.parquet, 2,539,608 rows)

2.1 Crosstab collision_energy_orig_units x ingest_lib

All 220,849 GNPS rows, all spectraverse "unknown" rows (47,075) and all drug_plus/masaryk rows have collision_energy_ev = None (319,865 rows with orig None; plus free-text orig such as "15, 30, 45, 60, 70 or 90 (nominal)" [416], "30.0,50.0,70.0" [304], "3 kV" [175] that were not parsed).

2.2 How orig becomes ev

Order gotcha: for pluskal_ms2 multi-energy rows the orig list keeps acquisition order ("[20.0, 15.0, 60.0, 40.0]", 16,762 rows; "[20.0, 30.0, 60.0, 40.0]", 20,978 rows) while collision_energy_ev is sorted ascending, so position i of orig does not correspond to position i of ev; one row has 3 orig values but 2 ev values. pluskal_ms2 multi-CE rows: 2 values 3,548; 3 values 82,949; 4 values 83,611; 5 values 4,598; 6 values 16 (these are MSnLib-style merged/stepped Orbitrap ID-X spectra).

2.3 enveda-180 (1,153,785 rows, 182,941 InChIKey14; 887,895 positive / 265,890 negative)

Merged rows (>= 2 energies) = 306,722 = 26.6%; the merged three-energy spectrum has roughly 2x the peaks of any single-energy spectrum (258 vs 88-132), and the two-energy merges are tiny (median 9 peaks) - almost certainly cases where one scan failed QC; the test set contains none of them. Grouping by (InChIKey14, adduct): 350,115 groups; 186,392 (53.2%) have all four spectra (20, 40, 60, merged); 73,755 have merged + 40 + 60 (no 20 eV single); 31,590 have only 40 + 60; 7,749 have the merged spectrum only; 46,578 have singles only. Original strings: "40" 244,197, "20,40,60" 240,127, "60" 229,217, "20" 173,189, "-40" 83,271, "-60" 70,248, "-60,-40,-20" 64,653, "-20" 46,941, "20,40" 1,037, "-40,-20" 739.

2.4 test.parquet (1,213 rows, all instrument_type = timsTOF, units always eV)

The visible test is 226 negative-mode rows (18.6%) and 987 positive. No 2-energy merges, no other energies.

2.5 What a merged enveda-180 spectrum physically is (code + numerical test)

Code (enveda/enveda-180, data_processing/extraction/extract.py lines ~330-452 and utils.py): for each adduct the m0-isotope MS/MS microscans of all precursor-selection events across >= 2 distinct collision energies are concatenated and centroided with centroid_tof_signal_df (consecutive TOF indices with gap < 2 are one peak; make_peak_centroid returns the intensity-weighted mean m/z and the sum of intensity); the record gets collision_energies = sorted(unique), mean_collision_energy = mean (= 40 for 20/40/60), collision_energies_merged = True, num_merged_microscans. Single-energy records are built the same way per (adduct, msms_id, collision_energy). MS1 intensity threshold 50 counts and "stray signal" filters (ms_prepper/filters.py: MIN_MS2_TOF_COUNT = 5) are applied upstream; there is no relative-intensity cutoff (minimum relative intensity in merged spectra: median 4.2e-5).

Numerical test on 600 random (InChIKey14, adduct) groups having all four spectra, reconstructing the merged spectrum from the three singles after de-normalising with base_peak_intensity (raw counts): the raw-count sum rule is the best fit in 595/600 (median per-peak MAE 8e-5 after re-normalisation; p90 1.2e-3), vs raw max 2, normalised sum 1, none 2. The merged peak set is a subset of the union of the singles in 334/400 (exact union only 10/400): peaks present in a single but absent from the merged spectrum have median relative intensity 5.9e-4 (p90 0.008, p99 0.08) - faint peaks are lost or re-clustered when the pooled TOF profiles are re-centroided. Base-peak raw counts per energy are very unequal: examples b20:b40:b60 = 399,341:341,591:66,697 and 130,526:23,809:3,704 - so the merged spectrum is a count-weighted sum dominated by the 20 and 40 eV scans. Consequence (Section 4 table): merged vs its own 40 eV scan cosine 0.874, vs 20 eV 0.858, vs 60 eV 0.693.

3. Merged multi-CE spectra across ecosystems

Enveda publications: the Bruker "Customer Insight" page says Enveda runs a timsTOF Pro 2 and relies on "clean MS/MS fragmentation patterns and ion mobility data" plus CCS, with no energies given; the PRISM foundation-model announcement (1.2 B spectra, spectral masking) gives no acquisition detail. Nothing published by Enveda that I could open states the merge rule - the code is the source of truth.

4. Head-to-head cross-platform comparisons (numbers)

Hoang et al. 2024, Anal Chem, DOI 10.1021/acs.analchem.3c05576 (PMC11007677). >100 standards on Agilent 6520/6546 QTOF, Bruker Impact II QTOF, Waters Synapt G2-Si, SCIEX TripleTOF 5600 and Thermo Q Exactive Plus at 0/10/20/40 eV (Thermo no 0 eV). Cross-platform cosine identification succeeded for 24 standards at 10 eV, 60 at 20 eV, 38 at 40 eV, <50 at 0 eV; examples at 20 eV: Z-FA-FMK 0.928-1.000, sphingosine 0.990-0.999, caffeine 0.982-0.994; "the average cosine scores did not necessarily increase with a greater number of fragments ... plateaued or slightly decreased". Recommendation: standardise at 20 eV. No timsTOF in the study.

Szabo, Schlosser, Vekey, Drahos, Revesz 2021, J Mass Spectrom 56:e4693, DOI 10.1002/jms.4693. Bruker Maxis II QTof CID vs Q Exactive Focus HCD (both N2): best-match energy pairs linear with slope 0.92 (R2 0.997); HCD a few eV lower; CID spectra contain 60% more b ions; survival yield 84-85% at the Orbitrap minimum (10 eV) ~ QTof 12 eV, QTof 98% at 6 eV and 99.4% at 0 eV; bimodal score-vs-energy curves for 60% of peptides on QTof vs 20% on Orbitrap.

Li et al. 2021, Nat Methods, DOI 10.1038/s41592-021-01331-z. Entropy similarity AUC 0.958 on 434,287 NIST20 spectra; "acceptable AUCs were even obtained when searching QTOF spectra against orbital ion trap MS/MS libraries or vice versa"; FDR < 10% at entropy similarity 0.75 on 37,299 natural-product spectra; gut-metabolome 770 compounds: 5.8% vs 9.6% FDR; median spectral entropy rises from 0.7-0.8 (< 15 eV) to 1.8-2.5 (> 45 eV).

No published small-molecule timsTOF-vs-Orbitrap MS2 similarity study was found (TIMS-Bench, Rajkumar et al. 2026, DOI 10.64898/2026.05.23.724673, benchmarks processing tools, not cross-instrument similarity). So I measured it in train.parquet: same InChIKey14 and same adduct ([M+H]+ or [M-H]-), precursor m/z within 0.02, precursor region (> m/z_p - 1) removed, sqrt-intensity cosine, 0.01 Da match tolerance (2,421 timsTOF/Orbitrap pairs; Orbitrap = pluskal_ms2, i.e. MSnLib Orbitrap ID-X flow injection; only 4 timsTOF/QTOF pairs qualified, so the QTOF column is from 52 spectra at 6 eV and is weak):

(medians; fraction >= 0.7: tims20/NCE20 0.68, tims40/NCE45 0.50, tims60/NCE60 0.34, merged/NCE45 0.48.) Read-outs: (i) a timsTOF 20 eV scan is an Orbitrap NCE 15-20 scan, i.e. 9-14 eV HCD at m/z 300-350 - the "same lab eV" assumption (NCE 33) is wrong by ~0.65x; timsTOF 40 eV ~ NCE 45 (~30 eV HCD at m/z 330, factor ~0.75); timsTOF 60 eV ~ NCE >= 60 (~40 eV HCD, factor <= 0.66, best match probably above the library's energy range). (ii) The merged spectrum behaves like a single mid-energy scan (best vs NCE 30-45) and is very close to its own 40 eV scan (0.874). (iii) Same-molecule, same-instrument, different-energy similarity (20 vs 60 eV: 0.285) is worse than cross-instrument at matched energy - energy mismatch, not vendor, is the dominant nuisance.

5. CID physics that is predictable with energy

• Survival-yield / breakdown curves. Precursor survival vs E_com is sigmoidal; CE50 (energy at 50% survival) is "easily determined, quantitative, highly reproducible, and could discriminate between structural and even positional isomers", "independent of cone potential and orthogonal to compound mass" over 54 compounds on a Waters QTOF (Kertesz, Hall, Hill, Grant 2009, JASMS, DOI 10.1016/j.jasms.2009.06.002). Over large mass ranges the energy for 50% fragmentation is linear in precursor mass, R2 > 0.996 (Li-PEG to 4.5 kDa) and R2 = 0.991 for multiply charged species across instruments - the degrees-of-freedom effect, consistent with fast energy randomisation/RRKM (Memboeuf et al. 2010, Anal Chem, DOI 10.1021/ac902463q). Converting CE50 to Ecom50 removes most of this: "The Ecom50 range was only 13.5% of the CE50 range" (Hall et al. 2014, Curr Comput Aided Drug Des, DOI 10.2174/1573409911666141231113516); Ecom50 is predictable from structure with r2 0.89-0.94 on 52 compounds (Hall et al. 2012, JCIM, DOI 10.1021/ci300092s) and was used by MolFind to remove ~28% of PubChem candidates per mass bin (Menikarachchi et al. 2012, Anal Chem, DOI 10.1021/ac302048x).
• Instrument transmission bias. RF-only ion-guide mass discrimination distorts relative abundances: "compounds that generate major product ions with m/z values < 150 have peak intensities that deviate substantially from their actual abundance" (Dubey et al. 2015, Metabolomics, DOI 10.1007/s11306-014-0732-0, PMC4422648; ten compounds 46-854 Da, oa-TOF).
• Energy-ordered pathways. Low E_com favours the lowest-barrier, entropically cheap channels: small-neutral losses (H2O, NH3, CO, CO2, HCOOH, CH3OH), glycosidic cleavages (-162/-146/-132 Da), retro-Diels-Alder and amide cleavages; these product ions are themselves intermediates and are consumed sequentially at higher E_com into aromatic/heteroaromatic core ions and low-mass "fingerprint" ions. In peptides this is the b-ion (low-E optimum) vs y-ion split that produces bimodal score-vs-energy curves (Szabo et al. 2021). Which specific neutral loss dominates is molecule-specific (that is what CFM-ID/ICEBERG learn); what is generic is the monotone shift of the intensity-weighted fragment m/z toward lower values and the depletion of the precursor.

Measured on enveda-180 single-energy [M+H]+ / [M-H]- spectra (300,000-row random sample; precursor = peaks within 10 ppm of precursor m/z; TIC = sum of all normalised intensities):

So for 300-500 Da protonated molecules on this timsTOF: at 20 eV the precursor survives in ~97% of spectra and carries ~11-13% of the TIC (base peak in 14%); at 40 eV it is gone in 63-72% (survival is the useful discriminator here - the mass dependence 72% -> 63% is the DOF effect); at 60 eV it is gone in 99%, and 76-94% of the fragment TIC is below m/z 150. Negative mode survives slightly better. The merged spectrum inherits the 20 eV precursor (base peak in 25-29% of merged spectra), which inflates cosine against any spectrum containing a precursor peak.

6. Collision-energy handling in in-silico fragmenters

7. Forced analogy: stellar spectroscopy (forward model x instrument response)

The analogy. A stellar classifier never compares an observed spectrum to a raw synthetic one; it computes F_obs(lambda) = F_model(T_eff, log g, [Fe/H]) * LSF x throughput(lambda) + noise, and fits with chi2 using a per-pixel noise model. The forward model is physics; the instrument response R (line-spread function, wavelength solution, throughput) is calibrated once per instrument from standard stars and applied to every template. Translated: a library or in-silico spectrum S_lib(f; E) is the "model atmosphere" output under condition E; a timsTOF-specific response R maps it into timsTOF-merged intensity space: S_tims(f) ~ R[S_lib(.; E_1..E_k); m_p, z, adduct]. Cosine on raw templates is the equivalent of comparing an unconvolved model to data - every "standard star" in Section 4 says the mismatch is large and systematic.

Concrete experiment (all ingredients exist in train.parquet).

1. Pairs. 2,523 InChIKey14 have both timsTOF and Orbitrap spectra (2,499 via pluskal_ms2, 1,040 also in GNPS, 253 riken, 238 mona, 210 massbank; 2,288 of them are enveda-180 molecules and 250 are enveda-np-examples, which matter because they are COCONUT-like); 3,468 (InChIKey14, adduct) timsTOF/Orbitrap pairs (2,198 [M+H]+, 996 [M-H]-, 167 [M+FA-H]-, 47 [M+Na]+) covering 10,939 timsTOF and 40,681 Orbitrap spectra; only 412 timsTOF/QTOF pairs (14,681 QTOF spectra). Split by Murcko scaffold of the InChIKey14, never by spectrum.
2. Response model R, three nested forms. (a) Energy-axis map only: library energies re-indexed by the Section 4 calibration (timsTOF eV -> NCE 20/45/60), then count-weighted sum w = (0.5, 0.4, 0.1) learned by non-negative least squares on the pairs - zero parameters beyond three weights and one scale factor. (b) Peak-wise transfer: for each fragment f present in any library energy, predict its timsTOF relative intensity from features [log I_lib(f) at each available NCE, m_f/m_p, m_f, m_p, adduct, polarity, is-precursor, is-neutral-loss-of-precursor] with a gradient-boosted regressor (also predicts presence/absence, i.e. the Dubey-type low-m/z transmission term). (c) Spectrum-conditioned: a small transformer taking the set of (m_f, I_lib at 3-5 NCEs) tokens and emitting timsTOF intensities, trained with a cosine/entropy loss - this is the "learned LSF".
3. Evaluation. Primary: cosine and entropy similarity of R(S_lib) to the real timsTOF merged spectrum on held-out scaffolds, versus the raw library spectrum (baseline from Section 4: median 0.55-0.69 depending on NCE). Decisive: a retrieval test that mimics the hidden test - for each held-out timsTOF merged spectrum, candidates = COCONUT + train structures within 10 ppm neutral mass, score each candidate by similarity between the timsTOF spectrum and R applied to the candidate's library spectrum (if any) or CFM-ID/ICEBERG prediction, report MRR@25 with vs without R. Secondary: apply R to CFM-ID 10/20/40 eV and ICEBERG predictions for the 250 enveda-np-examples molecules (natural-product chemistry, timsTOF) - the closest proxy to the hidden test.
4. Expected effect size. Form (a) alone should lift the merged-vs-Orbitrap median similarity from ~0.61 (NCE 20) toward the matched-energy ceiling (~0.70-0.82); if MRR@25 gains are < 0.01 on a timsTOF held-out set the analogy is not worth more effort.

Where it holds, partially holds, and breaks.

• Holds: the instrument response has large, systematic, learnable components - energy-axis offset (slope ~0.65-0.75 timsTOF-eV -> HCD-eV; Szabo's 0.92 QTof -> HCD), the count-weighted merge, low-m/z transmission (Dubey), and a precursor term. A response model is also exactly what Thermo's own NCE and Bruker's 1/K0 ramp are: crude R's.
• Partially holds: unlike an LSF, R is not shift-invariant or linear - the same lab energy deposits different internal energy in different molecules (DOF effect, R2 > 0.996 mass dependence) and ion-population statistics differ (Szabo: Orbitrap "larger number of collisions" and C-trap residence time; timsTOF adds TIMS accumulation and quad isolation of 2-3 Th). So R must take (m_p, adduct) and, in form (b)/(c), the fragment identity.
• Breaks: CID is not LTE - there is no temperature; the internal-energy distribution after multiple collisions is broad and kinetic, and fragment intensities are RRKM rate x observation-window products with kinetic shifts. A spectrum at one energy is not a "convolved" version of another; sequential fragmentation makes intermediate ions appear and disappear non-monotonically (b-ion vs y-ion bimodality). Hence R cannot be calibrated from "standard stars" and applied blindly: it is molecule-dependent, and the honest implementation is a learned conditional map (form c), i.e. the analogy collapses into "train a spectrum-to-spectrum translator", with the physics supplying the inductive biases (energy map, count weights, precursor survival, low-m/z bias). The chi2-with-noise-model half of the analogy does transfer cleanly: base_peak_intensity x normalised intensity recovers raw counts, so a Poisson-weighted similarity (down-weight peaks with < ~50 counts) is a free, physically justified improvement over cosine.

8. Gotchas / what could fool us

1. "eV" is not a common currency. Train stores Orbitrap NCE converted with (m/z)/500 and QTOF volts verbatim; empirically a timsTOF 20 eV spectrum matches Orbitrap NCE 20 (0.82) and only 0.32 vs NCE 60 - but the converted column says the NCE-20 spectrum of a 330 Da molecule is "13 eV" and the NCE-60 one "40 eV". Treating collision_energy_ev as comparable across ingest_lib will teach a model the wrong energy axis.
2. Sorted ev vs unsorted orig in pluskal_ms2 multi-energy rows: positional alignment is lost (and one row has mismatched lengths).
3. Ramps masquerade as merges: "Ramp 5-60 V" -> [5.0, 60.0] (1,397 rows), "Ramp 20%-70% (nominal)" similar; a 2-element list is not necessarily two scans.
4. 10,318 ion-trap NCE rows have ev = None, and 319,865 rows (all GNPS) have no energy at all - any CE feature must handle missingness explicitly, and missingness is library-identifying (GNPS/spectraverse/drug_plus).
5. The merged spectrum is count-weighted, so it is effectively a 20-40 eV spectrum with a 60 eV tail; it is not the equal-thirds Thermo stepped-NCE object, not a max-merge (mzmine SAME_ENERGY), and not a normalised average. Models trained on other "merged" conventions see a different object.
6. Merged num_peaks ~ 2x single (258 vs 88-132; p75 548), with many peaks < 1e-3 relative intensity; peak-count-sensitive scores change (Hoang: cosine plateaus/decreases with more fragments). Top-N truncation and the 0.003 cutoff used by ICEBERG change what a "merged" spectrum is.
7. The precursor survives at 20 eV (and therefore in every merged spectrum): base peak in 14% of 20 eV and 25-29% of merged [M+H]+ spectra. Any similarity that includes the precursor peak is inflated for every candidate with the right mass - strip the precursor and its isotopes (> m_p - 1.0) before scoring, as done in Section 4.
8. Instrument type is a chemistry label in train: all 1,154,969 timsTOF rows come from enveda-180 (Enamine REAL drug-like synthetics) and enveda-np-examples; a model can learn "timsTOF => Enamine-like" and then meet natural products on timsTOF in the hidden test. Validate on enveda-np-examples (1,184 rows, 99.6% in COCONUT) and on the 2,288 cross-instrument molecules, not on random enveda-180 rows.
9. Visible test rows are verbatim enveda-180 rows - any local score on them is leakage; also hold out by InChIKey14 and by CE pattern so the 20/40/60/merged quartet of one molecule never spans the split.
10. [2M+H]+ / [2M+Na]+ in positive mode dissociate to monomer ions at 20 eV; mass-binned survival statistics computed over all adducts show "72% precursor absent at 20 eV for m/z 500-700" purely from dimers (I made this mistake before restricting to [M+H]+/[M-H]-).
11. Two-energy merges in train (1,942 rows) are degenerate (median 5-9 peaks) and absent from test - drop them from training or they teach "merged = tiny".
12. NCE equivalents of timsTOF 60 eV exceed typical library energies: for a 300 Da ion, 60 eV lab is NCE 100 by formula and ~NCE 60-70 by the empirical calibration; NIST20/MSnLib rarely go that high, so 60 eV singles are the hardest to match against libraries and the merged spectrum's 60 eV component is largely unmatched.
13. The enveda-180 preprint DOI in the README did not resolve for me; acquisition details (isolation width, microscans, separate injections vs CE stepping) are unverified - treat any claim about them as low confidence.

9. Implications for the challenge

1. Calibrate the energy axis before any library or in-silico comparison: use timsTOF 20 eV ~ NCE 15-20, 40 eV ~ NCE 45, 60 eV ~ NCE >= 60 (Section 4), i.e. NCE_equiv ~ 0.65-0.75 x eV x 500/(m/z) for z = 1; feed MassFormer/FraGNNet/ICEBERG-2 these NCEs, and CFM-ID 10-20 / 20-40 / 40 eV.
2. Synthesise the merged object exactly: for any candidate spectrum set (predicted or library), build the count-weighted sum (start with weights 0.5/0.4/0.1 from the base-peak ratios, then fit by NNLS on the 186,392 all-four groups), re-normalise to base peak, drop peaks < 1e-3; compare merged test spectra to that, and single-energy test spectra to the matched energy.
3. Train on singles, not only on merged: 73% of train rows and 75% of test rows are single-energy; a spectrum-to-fingerprint model should take the CE pattern as a conditioning token (20 / 40 / 60 / merged) and the model can be made invariant by augmenting - synthesise merged spectra from the triplets and single spectra from merged minus the others where available.
4. Strip precursor and isotopes, and use Poisson-weighted similarity: recover counts with base_peak_intensity; weight peaks by sqrt(count) and ignore peaks < ~50 counts; evaluate entropy similarity (Li 2021) in addition to cosine.
5. Use precursor survival and the intensity-weighted fragment-m/z ratio as cheap, energy-explicit features: at 40 eV, survival (present in 28-37% of [M+H]+ spectra at 250-450 Da) carries Ecom50-type stability information that discriminates isomers (Kertesz 2009); a predicted Ecom50 (Hall 2012 QSAR, r2 ~0.9) is a candidate-ranking feature orthogonal to fingerprints.
6. Run the response-model experiment of Section 7 with form (a) first (one afternoon: energy map + three weights), measure MRR@25 on a scaffold-held-out timsTOF set with a COCONUT +/- 10 ppm pool; proceed to forms (b)/(c) only if the gain is >= 0.01.
7. For CFM-ID/ICEBERG candidate scoring on the hidden test, score each candidate against each single-energy spectrum at its matched prediction energy and combine per molecule (median 3 spectra per molecule); do not score only the merged spectrum.
8. Build the local validation set from enveda-np-examples (1,184 rows) plus the 2,288 cross-instrument enveda-180 molecules, scaffold-split, with the hidden-test CE mix (19/28/26/27% for 20/40/60/merged in train; 16/30/30/25% in visible test).

10. Open questions

• What exactly did Enveda acquire: isolation width, TIMS ramp range, number of microscans per energy, separate injections per CE vs in-run stepping, and whether the hidden test used the identical pipeline (the merge rule is inferred from the public code and a 600-group fit, not from a statement).
• Does the enveda-180 preprint (ChemRxiv DOI as cited in the README) give a per-energy QC filter that explains the 73,755 groups lacking a 20 eV single scan?
• Energy range of MSnLib/NIST20 Orbitrap spectra above NCE 60: is a timsTOF 60 eV scan ever matchable to a library energy?
• Does ICEBERG 2.0 accept an arbitrary NCE, and what did it see for merged training spectra?
• Is the Enveda PRISM model (1.2 B spectra) trained on merged or single-energy timsTOF spectra? Unknown; nothing published.
• Would a Poisson-weighted / count-aware similarity beat sqrt-cosine on the 2,421 cross-instrument pairs? (One-hour test with the tables in the JSON file.)

Sources

• Train/test parquet analyses: C:\Temp\casmi26\data\train.parquet (2,539,608 rows) and test.parquet (1,213 rows), pyarrow 25; scripts in the session scratchpad (ce_analysis.py ... ce_analysis5.py); all tables also in cid_fragmentation_physics_collision_energy.json.
• enveda/enveda-180 GitHub repository (README; data_processing/extraction/extract.py, utils.py, centroid.py; ms_prepper/reader.py, filters.py), https://github.com/enveda/enveda-180 (cloned 2026-10-07); README cites Krettler et al. 2026 ChemRxiv DOI 10.26434/chemrxiv.15004319/v1 (not resolved - unverified).
• Enveda GitHub organisation listing https://github.com/orgs/enveda/repositories; Bruker "Customer Insight Enveda" https://www.bruker.com/en/landingpages/bdal/customer-insight-enveda.html; TIMS-Bench README (Rajkumar et al. 2026, DOI 10.64898/2026.05.23.724673).
• Wang A. et al. 2024, Microbiome Research Reports, DOI 10.20517/mrr.2024.21 (PMC11480776).
• Meier F. et al. 2018, Mol Cell Proteomics, DOI 10.1074/mcp.TIR118.000900 (PMC6283298).
• Vasilopoulou C.G. et al. 2020, Nat Commun, DOI 10.1038/s41467-019-14044-x (PMC6965134).
• Monteiro F.S., Buzatto A.Z. 2026, STAR Protocols, DOI 10.1016/j.xpro.2026.104607 (PMC13251676).
• Thermo Fisher PSB104 "Normalized Collision Energy Technology", https://documents.thermofisher.com/TFS-Assets/CMD/Warranties/PSB104-Normalized-Collision-Energy-Technology-EN.pdf.
• Szabo D., Schlosser G., Vekey K., Drahos L., Revesz A. 2021, J Mass Spectrom 56:e4693, DOI 10.1002/jms.4693 (author PDF at pak.elte.hu).
• Hoang C. et al. 2024, Anal Chem, DOI 10.1021/acs.analchem.3c05576 (PMID 38529642, PMC11007677).
• Li Y. et al. 2021, Nat Methods, DOI 10.1038/s41592-021-01331-z (PMC11492813).
• Diedrich J.K., Pinto A.F.M., Yates J.R. 2013, J Am Soc Mass Spectrom, DOI 10.1007/s13361-013-0709-7 (PMC3815594).
• Horai H. et al. 2010, J Mass Spectrom, DOI 10.1002/jms.1777 (Europe PMC record).
• Brungs C. et al. 2025, Nat Methods, DOI 10.1038/s41592-025-02813-0 (PMC12510872); MSnLib Zenodo record 11163380.
• Goldman S. et al., ICEBERG, arXiv 2304.13136 (PDF); samgoldman97/ms-pred README.
• Bushuiev R. et al., MassSpecGym, arXiv 2410.23326 (HTML v3).
• Young A. et al., MassFormer, arXiv 2111.04824; Young A. et al., FraGNNet, arXiv 2404.02360 (HTML v2).
• Wang F. et al. 2022, CFM-ID 4.0 web server, Nucleic Acids Res, DOI 10.1093/nar/gkac383 (PMC9252813); Wang F. et al. 2021, Anal Chem, DOI 10.1021/acs.analchem.1c01465 (cited via the NAR paper; not opened directly).
• Kertesz T.M., Hall L.H., Hill D.W., Grant D.F. 2009, J Am Soc Mass Spectrom, DOI 10.1016/j.jasms.2009.06.002.
• Memboeuf A. et al. 2010, Anal Chem, DOI 10.1021/ac902463q.
• Hall L.M. et al. 2012, J Chem Inf Model, DOI 10.1021/ci300092s; Hall L.H. et al. 2014, Curr Comput Aided Drug Des, DOI 10.2174/1573409911666141231113516; Menikarachchi L.C. et al. 2012, Anal Chem, DOI 10.1021/ac302048x; Dubey R. et al. 2015, Metabolomics, DOI 10.1007/s11306-014-0732-0 (PMC4422648).
• Kaggle data-page facts (collision_energy_ev list semantics; enveda-180 1,153,785 spectra / 182,941 structures) taken from GitHub mirrors/write-ups surfaced by search (eganeganegan/casmi26 and others; the Kaggle page itself is JS-rendered and was not read).