About This Tool
The Atlas of Proteomic Technologies (APT) is built and maintained by Ignition Scientific to help researchers select the right proteomics platform for their study. It synthesizes published head-to-head comparison data with expert assessment of platform capabilities from scientists at Oxford University and the Royal College of Surgeons in Ireland. The tool is updated with each release to reflect new evidence and platform changes. Scores and recommendations reflect the authors' assessment based on published literature and are intended to guide study design decisions, not to serve as definitive or commercial platform rankings.
How to Cite
If you use this tool to inform a study design or publication, please cite the accompanying preprint:
This preprint has not yet been peer reviewed; the citation will be updated when a peer-reviewed version is available.
Code and data: see Reproducibility
How This Tool Was Built
The scoring framework and recommendation engine were developed iteratively, combining published evidence, expert judgement, and real-world researcher feedback.
Curate evidence base
10 independent head-to-head studies collected, covering the majority of evaluated platforms across plasma, serum, CSF, and other matrices.
Model draft
On 30 March 2026, Whelan used Claude (Anthropic) to draft scores for the six platforms on nine analytical dimensions. The prompt file generated in the same session as the draft also told the build to adjust scores so that platform totals fell between 28 and 33; the authors did not set this range or use platform totals when scoring or adjudicating.
Author scoring & adjudication
↩ iterative revision returns hereWhelan & Smith-Byrne scored the six platforms on the nine dimensions before opening the model's draft, then compared the two sets and adjudicated. Before release on 2 April 2026, 19 of the 54 draft scores changed (18 by one point; 11 up, 8 down) and a tenth dimension, pQTL accuracy, was added. The authors' pre-adjudication set was not retained. Scientific justification documented for every score. Fairness simulation run across 1M+ weight combinations to identify and correct structural biases in the recommendation engine.
Public beta release
Recommendation engine and scoring framework published online. Researchers invited to complete the tool and submit feedback via embedded survey.
Real-world calibration
65 study profiles collected from researchers in academic, industry, clinical and other settings in eight countries; 63 could be scored, since two required systematic PTM detection, which excludes every platform. Responses were used to calibrate recommendations and to refine scores and the algorithm where real use departed from the simulation.
This tool is a starting point, not a final verdict
Scores and recommendations are based on published evidence and expert assessment at the time of release. Before committing to a platform, consult directly with each vendor's scientific and commercial teams, request a pilot study where possible, and review the primary literature for your specific sample type and research question. Platform performance varies substantially by study design, sample handling, and protein targets of interest.
Scores are relative, not absolute
APT scores reflect each platform's standing within the current evaluated set, not a verdict on any platform's absolute quality. A score of 3/5 does not mean a platform is mediocre; it means that, on that dimension, other platforms in this comparison currently perform better or worse. All six evaluated platforms are commercially viable, actively used in peer-reviewed research, and appropriate choices for the right study design.
Scores are updated as evidence accumulates. Platforms with limited published validation data are scored conservatively; scores will be revised, in either direction, as independent head-to-head studies emerge. The version history below documents every score change and its scientific basis. If you believe a score is inconsistent with the published evidence, contact us.
Independence, trademarks, and terms of use
The Atlas of Proteomic Technologies is an independent, educational resource provided free of charge. It is not sponsored by, affiliated with, or endorsed by any platform or vendor named, and Ignition Scientific has no financial relationship with the companies whose products are compared. Scores and recommendations are the authors' good-faith interpretation of publicly available data and expert judgment as of the date shown; they are offered as opinion, not statements of fact, and are not definitive or commercial rankings, nor procurement, legal, or investment advice.
Product and company names (including Olink, SomaScan, SomaSeq, Alamar NULISA, Nomic nELISA, Seer Proteograph, and Biognosys TrueDiscovery) are trademarks of their respective owners and are used here for identification and comparison only. The tool is provided "as is," without warranty of any kind; verify all specifications and pricing with the vendor before making purchasing decisions. If you believe any statement is inaccurate or out of date, email chris@ignitionscientific.com and we will review and correct it promptly.
Author Contributions
Christopher D. Whelan conceived and developed the Atlas of Proteomic Technologies, including the scoring framework, recommendation engine, evidence curation, and web application. Karl Smith-Byrne provided scientific oversight and critical review of the scoring methodology. C.D.W. and K.S-B. contacted their respective industry and academic networks to gather early feedback and to inform future iterations of the tool.
Platform Scoring Methodology
Each platform is scored on 10 dimensions using a 1–5 scale. Scores reflect a combination of published evidence and expert assessment. The basis for each dimension is indicated below. Scores are reviewed and updated as new comparative data is published.
Breadth of proteins measured
Based on vendor-reported assay counts and published panel comparisons. Reflects the number of unique protein targets with validated assays. Scores follow a protein-count rubric: <1,000 proteins = 1/5; 1,000–3,000 = 2/5; 3,000–5,000 = 3/5; 5,000–7,000 = 4/5; >7,000 = 5/5. Applied to evaluated platforms: Nomic nELISA (~1,100 proteins) = 2/5; Olink Explore HT (~5,400 proteins) = 4/5; Biognosys TrueDiscovery (~4,200 proteins in plasma, up to ~13,800 in tissue) = 4/5 as a cross-matrix middle estimate.
1–5 scale
Reproducibility / CV
Based on median CV reported in Kirsher et al. (2025) Alkahest benchmarking study, a direct comparison of eight platforms on plasma from the same donors.
1–5 scale
Target confidence, cis-pQTL support, epitope risk
Based on cis-pQTL enrichment rates from Eldjarn et al. (2023) and cross-platform correlation studies. Dual-antibody and pre-colocalized antibody platforms score highest. SomaSeq scores 2/5, reflecting the lowest cis-pQTL support (~43%) of any evaluated platform and elevated single-aptamer epitope-binding artifact risk. A score of 2/5 (rather than 1/5) reflects that SomaSeq detects real protein signals; the 1/5 score was revised upward to avoid an artefactual priority-miss penalty that suppressed SomaSeq recommendations in broad discovery scenarios where specificity is not the limiting factor. Two caveats apply symmetrically: a cis-pQTL can itself arise from an epitope effect, in which the variant alters reagent binding rather than protein abundance, so cis-pQTL support is an imperfect specificity proxy on any affinity platform; and mass-spectrometry proteogenomics has confirmed correct targeting for a subset of affinity assays, indicating the issue is assay-specific rather than uniform.
1–5 scale
Ability to detect low-abundance proteins
Evidence-based for established platforms (LOD comparisons, below-LOD rates in plasma). Based on Kirsher et al. (2025): SomaSeq scores 4/5, reflecting 96.2% of proteins detected across all samples and 97% of assays within linear range: broad consistent detection comparable to Olink. Note that aptamer cross-reactivity means some SomaScan signal may reflect off-target binding rather than genuine protein detection; the 2/5 specificity score captures this separately. Nomic nELISA scores 3/5: as a sandwich ELISA format, it does not benefit from the signal amplification of Olink's proximity extension assay (PEA), and there is insufficient published head-to-head LOD evidence to justify a score of 4/5 at this time. Expert assessment for other newer platforms with limited published LOD data.
1–5 scale
Lower cost per sample = higher score
Scores are commercial price bands for the standard discovery service at volume (orders of roughly 1,000 samples or more), drawn from 2024 to 2026 quotations and published rates: 5 = under about $75 per sample; 4 = about $75 to $250; 3 = about $250 to $1,000; 2 = about $1,000 to $3,000; 1 = above about $3,000. The bands are approximately logarithmic, and their edges sit between observed service prices rather than at exact thresholds, so a platform's band does not change if an edge shifts within the gap around it. Smaller batches price higher on every platform, and population-scale consortium pricing is excluded because it is not available to individual studies. Cost structures evolve; consult vendors for study-specific quotes.
1–5 scale
Samples per day/week
Based on vendor-reported sample throughput and published workflow descriptions. Reflects steady-state capacity, not maximum theoretical throughput.
1–5 scale
Type of quantification: 4 = physical concentration (pg/mL); 3 = proteome-wide comparable label-free MS; 2 = within-assay relative units
Classifies the type of quantification each scored (standard discovery) product reports. 4 = physical concentration in mass units (ELISA-calibrated pg/mL against recombinant standard curves): Nomic nELISA, the only scored product reporting absolute concentrations. 3 = proteome-wide comparable relative quantification, i.e. label-free MS intensity, on a common scale that supports cross-protein abundance comparison (iBAQ / Total Protein Approach) but is not a physical concentration: Seer Proteograph XT and Biognosys TrueDiscovery. 2 = within-assay relative units only, not comparable across targets (Olink NPX, SomaScan RFU / SomaSeq read counts, NULISA NPQ): Olink, SomaSeq, Alamar NULISA. A score of 5 would require MS absolute quantification with isotope-labelled internal standards (SIS peptides), which no scored panel provides as standard. Note that several vendors offer absolute quantification (pg/mL) via separate products not scored here: Olink (Flex / Target 48 panels with per-assay calibrators), Alamar (NULISAseq AQ panels), and Biognosys (TrueSignature targeted MS).
1–5 scale
Breadth of sample matrices with published validation (plasma, serum, CSF, tissue, cell lysate)
Reflects the breadth of sample matrices with published validation, not merely claimed compatibility. Biognosys TrueDiscovery scores 5: mass-spectrometry digestion is matrix-agnostic and yields label-free intensities on a common scale, so a single workflow is validated across plasma, serum, CSF, solid tissue and cell lysate. Seer Proteograph XT scores 3 despite also being MS: its nanoparticle enrichment requires liquid input, so the scored XT product covers biofluids only (solid tissue requires the separate Proteograph DIRECT product). The affinity platforms (Olink, SomaSeq, NULISA) score 3: their deep published validation is concentrated in plasma, serum and CSF. SomaScan/SomaSeq offers the broadest sample-matrix kit range of any affinity platform (plasma, serum, urine, CSF, tissue homogenate, cell lysate), but aptamer signal is matrix-dependent and is not comparable across matrices, so each matrix is validated in isolation rather than as a unified cross-matrix workflow; kit availability is not the same as published cross-matrix validation depth, and the bulk of published SomaScan evidence remains plasma/serum. A higher score would also overstate its fit in matrices where it is not the established choice (e.g. cell-culture secretome, led by Nomic). Nomic nELISA scores 2: its published validation is concentrated in cell-culture/secretome, with emerging blood data and no solid-tissue validation.
1–5 scale
Published head-to-head data available
Reflects the volume and quality of published head-to-head comparative studies at time of last update. Platforms with UKB-PPP or ARIC cohort data score highest.
1–5 scale
Reliability for pQTL studies; lower epitope-binding artifact risk
Scores reflect demonstrated performance in published pQTL studies, not theoretical capability. MS-based platforms are theoretically free of epitope-binding artifacts, but scores are assigned based on the volume and consistency of published pQTL validation data available at the time of this release. Antibody and aptamer-based platforms are both susceptible to epitope-binding artifacts; Olink's dual-antibody PEA design may reduce single-epitope artifact risk but does not eliminate it. SomaSeq scores 2/5: aptamer-based assays are susceptible to binding QTLs (genetic variants affecting aptamer affinity rather than true protein abundance), which reduces confidence in pQTL signal interpretation relative to antibody-based platforms. NULISA scores 3/5, reflecting limited published pQTL validation data despite strong assay specificity; this score should be interpreted cautiously for genetics-first study designs. Biognosys TrueDiscovery also scores 3/5 despite being MS-based, reflecting the near-absence of published pQTL studies using that platform to date; its score may be revised upward as the field accumulates MS-based pQTL data.
1–5 scale
Platform Catalog and Integration-Readiness
The All Platforms tab situates the six scored platforms within the broader landscape. It catalogs the protein measurement platforms known to us at the time of writing, where a platform measures proteins as its primary readout by any modality (affinity or aptamer multiplex, immunoassay, mass spectrometry, or single-molecule protein sequencing) and is a distinct commercial or pre-commercial offering. It excludes extracellular-vesicle and particle-enrichment chemistries, spatial and tissue-imaging platforms, and non-proteomic comparators such as seed-amplification assays. Each entry records an explicit availability status, so superseded products (for example Olink Explore 3072) are documented with their successor rather than dropped.
Entries are one of three kinds. Services and assays are the unit a laboratory procures, and are the only entities ranked. Instrument lines (for example Orbitrap Astral, Bruker timsTOF) are shown as the enabling hardware that services run on, and sample-preparation chemistries are shown the same way. Because a service such as Seer Proteograph or Biognosys TrueDiscovery is defined by the mass spectrometer it runs on, ranking the instrument separately would double-count the same measurement; instruments and chemistries therefore sit in an enabling layer, and each service records the instrument it uses.
Platforms are scored only where the evidence supports it, at three levels. Platforms characterisable on all ten comparison axes receive the full score and form the integrated tier; a targeted or configurable panel with no fixed large content list, including any assay measuring fewer than 100 proteins, is treated as a targeted panel: not eligible for the integrated tier and not ranked on integration-readiness, regardless of how well it is otherwise characterised, since a platform that is not a candidate for integration has no meaningful position on a readiness-toward-integration scale. Such panels are catalogued as a separate group. Other deployable platforms whose measurement chemistry carries independent, peer-reviewed validation receive a lighter integration-readiness score and are ranked. Pre-commercial platforms and those without independent validation are catalogued as emerging, with no readiness score and no rank, since a number would imply a discrimination the evidence cannot support. An independence gate governs the readiness score: vendor posters, press releases, and specification sheets do not qualify. Under this gate a newly launched platform whose only evidence was a vendor conference poster (ProteinXI) is catalogued as emerging despite its fixed assay list, while platforms built on independently validated chemistries are scored even where a specific product's performance figures are vendor-stated.
The integration-readiness index combines three components reported for every entry: an integration tier (whether the platform exposes a fixed published per-protein list, a defined or configurable panel, or no fixed analyte list), commercial maturity on a four-level ladder, and validation depth indexed to independent peer-reviewed use in human cohorts. The composite is 0.5 x maturity plus 0.3 x validation plus 0.2 x (4 minus tier), with maturity and validation scored 0 to 3 and tier numbered 1 to 3, rounded to the nearest 0.25. It separates kind from rung: a mature open-discovery mass-spectrometry platform is labelled a reference method, not penalised as immature. Integrated platforms are held to the same evidence standard; where a headline specification is vendor-stated, such as the roughly 1,000-plex coverage of the Nomic Omni 1000 whose independent validation extends to about 191-plex, that is reflected in a low evidence-depth score rather than exempted. The catalog comprises 46 platforms: six are APT-assessed on all ten axes (eight catalog rows, since the SomaScan lineage spans array and NGS readouts and NULISA spans two integrated panels); nine carry an integration-readiness score as candidates for a future version; eight are targeted panels; eight are early or pre-commercial technologies; two are retired; and eleven are enabling instrument lines or sample-preparation chemistries. Thirty-eight are currently deployable, verified against vendor sources for 2026, including the consolidation of the SomaScan menu onto Illumina SomaScan Discovery (array) and Illumina SomaSeq Discovery (NGS) after Illumina's acquisition of SomaLogic.
Evidence Curation
A study counts as an independent head-to-head study in the Evidence Base if it meets all of the following criteria:
- Peer-reviewed publication or preprint from a recognized repository (bioRxiv, medRxiv)
- Head-to-head comparison of two or more high-plex proteomics platforms on the same biological samples
- Not vendor-authored: no author is employed by the vendor of a compared platform
- Quantitative comparison of at least one analytical metric (CV, correlation, pQTL enrichment, or disease association)
- Plasma, serum, or other biofluid sample type (tissue-only studies excluded from primary curation)
Other papers cited on this site are listed as foundational rows and are not counted. Vendor authorship is marked on the row.
Recommendation Engine and Fairness Testing
The recommendation engine computes a weighted sum of platform scores across all dimensions. Users set importance weights (1–5) for seven dimensions via sliders, all defaulting to 3. Additional adjustments are applied automatically based on study context: large cohort sizes add a modest boost to cost and throughput weights (xlarge +1, large +0.5; halved in v3.0, see Calibration process below); CSF sample type amplifies the sensitivity weight; a CNS biomarker toggle, covering neuroscience applications and pTau species detection (pTau-181, pTau-217, pTau-231), elevates sensitivity (by 6 points for non-CSF matrices, 8 for CSF) and specificity (by 2 points for non-CSF, 4 for CSF), with the stronger CSF override reflecting the clinical pTau measurement context; and a longitudinal/repeat sampling toggle elevates precision by 2 points, reflecting the importance of within-person change detection in repeat-measurement studies. Multi-matrix validation carries a fixed context weight that scales with how demanding the selected matrix is. A soft penalty (0.85×) applies when a platform scores 1/5 on a dimension the user has personally set to 5 (v3.0: gated on the user's own raw slider, not the context-boosted effective weight). Two hard filters (PTM detection mode and absolute quantification requirement) remove ineligible platforms from ranking rather than merely penalising them. Evidence depth (how much published head-to-head data exists for a platform) is not part of this weighted sum at all: it was removed from Help Me Choose's composite score at v2.7, on the grounds that confidence in a measurement is not an attribute of the measurement itself, and including it systematically penalised newer platforms for commercial immaturity rather than technical deficiency. It remains visible as a standalone reference dimension elsewhere in the Atlas (Platform Comparison, Evidence Base) but plays no role in any Help Me Choose recommendation. In Help Me Combine, evidence depth is one of eight selectable measurement properties but defaults to a weight of zero, so it is likewise excluded unless a user explicitly opts in via "Customize measurement properties." Platforms with an evidence depth of 2 or less are flagged as lower confidence in both engines' results; the flag does not change any score or recommendation.
Systematic bias simulation
To assess whether the engine could unfairly favour or suppress any platform, we run a full deterministic enumeration of 1,259,712 scenario-weight combinations: every combination of three weight levels (low, medium, high) across all seven user-adjustable dimensions, crossed with six research goals, six sample types, four study sizes, and four toggle-state combinations (CNS-focus × longitudinal). This is a complete enumeration, not a sample, so it carries no seed and reproduces exactly. Each goal, sample type, study size, and toggle-state combination is weighted by its observed real-world prevalence among 65 real Help Me Choose respondents rather than treated as equally likely; see Baseline weighting methodology, below, for why. For each scenario, every eligible platform within 3 points of the top score is identified (its "tie group" - the same condition HelpMeChoose.jsx uses internally to decide whether to show a single winner or a tied set). Each scenario resolves to exactly one of two mutually exclusive outcomes for a given platform: co-recommended - the tie group has two or more members and this platform is one of them, so the app shows "N platforms are statistically tied" for all of them and singles out no winner - or outright win - the tie group has exactly one member, this platform, and no other platform is even within 3 points; this is the only condition under which the app declares a single "Best overall match" rather than a tie.
The tables below lead with co-recommended rate. Calibrated match percentages are typically only a point or two apart (v2.3), so a statistically tied field is the more common outcome and the one the live app itself has shown users, via the tie band, since that release; leading with outright win rate would overstate the precision of a comparison the underlying scores do not support to that degree. Outright win rate is still reported in full for every dataset, not omitted - it answers a real, different question (is there a single best choice, with nothing else close) - but it is the minority outcome even for the platform that leads on it: no platform stands alone in 62.9% of baseline scenarios, 62.1% of the secondary-check slice below, or 65.1% of scored real-world respondents. That imbalance is a property of the scenario population, not of any one platform; the full breakdown by tie-group size is in the downloadable reproducibility bundle rather than reproduced in full here. Neither column sums to 100% across platforms, for opposite reasons: co-recommended rates can sum to more than 100%, since every platform in a multi-way tie is counted once for that one scenario; outright win rates sum to less than 100%, for the reason above.
Baseline weighting methodology
Earlier versions of this simulation (and an earlier manuscript draft) enumerated every goal, sample type, study size, and toggle combination as equally likely and called that the "baseline." That is not a neutral choice: real usage is far from uniform. We had already rejected equal weighting for the two toggles specifically, since it would imply a CNS-focus rate of 50% - flagged as unrepresentative before the true rate was known. The same problem existed for goals, sample types, and study sizes, just previously unaddressed: pharmacoproteomics, for example, was 0% of real sessions but received full weight in the old grid, while plasma (55.4% of real sessions) and discovery (55.4%) each received only a sixth. The baseline below now reweights all four axes by observed prevalence among the 65 real respondents, so it answers "what does a randomly selected real user experience" rather than "what happens if every theoretical combination is equally likely." One platform shows why this matters. Under the old uniform toggle weighting, Alamar NULISA's simulated outright win rate reached 23.7%, seeming to lead the field, only because CNS-focus (where NULISA is strongest) was treated as half of all scenarios rather than its true rate. The field's top win rate barely moved when the toggles were first reweighted, because a toggle-insensitive platform held first place throughout; the swing showed up only when every platform was checked, not just the leader. Reweighting goals, sample types, and study sizes on the same principle brings NULISA's simulated outright rate to 4.7%. The previous uniform methodology is retained, unchanged, in the downloadable reproducibility bundle as a clearly labelled superseded reference (not deleted), so any previously published number under it can still be reproduced exactly. This reweighting affects only how the aggregate simulation is summarized, not any individual real user's recommendation, which already reflects that user's own stated situation. It does introduce a real limitation the uniform grid did not have: the baseline now depends on the 65-respondent sample being reasonably representative of true usage, worth refreshing as more real usage data accumulates, the same as evidence-base scores already are.
Calibration process
Initial simulation revealed structural biases including one platform winning more than 50% of all scenarios at equal default weights, another platform never appearing in the top three under any default-weight scenario, and a fixed context multiplier for tissue and cell lysate matrices that disproportionately favoured a single platform. Each identified bias was reviewed against the underlying scores: where a bias reflected an inaccurate score, the score was revised with scientific justification; where it reflected a genuine platform characteristic, it was retained.
A v3.0 re-validation of this simulation against the current, unchanged score matrix identified corrections in three areas, none involving any platform's score. The first concerns how user inputs translate into effective weights. The soft priority-miss penalty was gated on the context-boosted effective weight rather than the user's own slider, so an automatic goal adjustment could trigger a penalty on a dimension the user never touched; it now gates on the raw slider value. Relatedly, the large/extra-large study-size boost applied the same input to both the cost and throughput weights at full strength, double-stacking one input onto two dimensions that happen to be Nomic nELISA's strongest, tied-max scores; the boost is now halved. The second area is matrix weighting: the multi-matrix-validation weight for tissue samples (8) was disproportionate against the same weight for the other two atypical matrices, cell culture and multiple sample types (5 each), with no documented reason tissue should carry more: at 8, a platform with the maximum flexibility score earned 40 points from matrix compatibility alone, more than the maximum possible contribution from a fully-maxed priority slider (5 weight × 5 score = 25), letting sample type override any explicit user priority regardless of how strongly stated. Tissue now matches the other two atypical matrices at 5. The per-platform tissue scores reflect real validation differences: Biognosys TrueDiscovery handles native solid tissue and homogenates directly (5/5); Olink Explore HT and SomaSeq are validated in tissue homogenates, the normal affinity-platform tissue workflow, and score 3/5, the same as their general multi-matrix validation (the earlier tissue-specific downgrade to 2/5 was removed in v3.5, see version history); and Nomic nELISA, Alamar NULISA, and Seer Proteograph XT have no published solid-tissue validation and score 1/5. A separate v3.0 correction closed a related gap: a platform with no validation for the selected matrix could still win outright on unrelated strength, so the soft priority-miss penalty (0.85×, described above for slider dimensions) now also fires when a platform's multi-matrix validation score for the chosen matrix is 1/5. That brings the tissue outright rate of the unvalidated platforms (Nomic, NULISA, Seer) to 0%, so cost or throughput strength can no longer buy back a recommendation for a matrix the platform has never been deployed on. From v2.5 onward, the win-share figures reported as this analysis's current output are reproducible exactly from the archived engine specification and sweep script. This does not extend to prior-methodology figures quoted elsewhere for before/after comparison (for example in the version history below): those were not version-pinned to an archived, independently runnable state and are not claimed as reproducible.
A separate correction was needed in how the simulation tallies a "win," not in the recommendation engine, which was unaffected throughout. The live app declares a sole winner only when no other platform at all is within the 3-point tie band; a positive but sub-3-point margin is still shown to users as "N platforms are statistically tied," not a win. Two earlier tallying methods (crediting whichever platform had the single highest displayPct, then a revision giving fractional credit for exact numeric ties) both missed this: a platform with the numerically highest score could still be credited with an outright win in a scenario the app itself displays as tied, whenever its margin over the next platform was positive but under 3 points. For example, displayPct values of 68/66/65/65/61/54 put four platforms within 3 points of the 68, and the app shows all four as tied, but both earlier methods credited the 68 alone with a full win. Outright win is now defined identically to the app's own condition (exactly one platform in the tie group). A further revision separated this from co-recommended rate: an intermediate version reported a single "top-tier rate" that combined a platform's own outright wins with its co-recommendations, and listed the resulting "no sole winner" share as an additional row inside the platform table. The current version reports outright win and co-recommended as separate, mutually exclusive columns, with the tie-group-size distribution reported on its own as a scenario-level statistic. Every superseded definition is documented, not deleted, in the downloadable reproducibility bundle. Olink Explore HT's precision score (2/5, its weakest scored dimension) is independently evidenced by two published CV comparisons (Kirsher et al. 2025: 26.8%; Rooney et al. 2025 ARIC cohort: 35.7%) and was not revised during any of this.
Fairness simulation results: baseline (prevalence-weighted)
| Platform | Co-recommended | Outright win | Typically competitive when |
|---|---|---|---|
| Seer Proteograph XT | 39.2% | 12.3% | Unbiased deep discovery or pQTL studies in plasma |
| SomaSeq | 33.4% | 9.4% | Coverage or precision prioritised; broad discovery designs |
| Olink Explore HT | 33.2% | 3.1% | Throughput prioritised (tied with Nomic); broadly competitive across most goals |
| Nomic nELISA | 28.3% | 6.7% | Low cost per sample, throughput (tied with Olink), cell-culture/screening work, or absolute quantification required |
| Alamar NULISA | 17.4% | 4.7% | CNS biomarker focus, CSF matrix, or ultra-sensitivity required |
| Biognosys TrueDiscovery | 16.5% | 1.0% | Tissue and cost-sensitive studies; still lowest overall, and real-world sessions rarely specify tissue (~6%) |
The table reports the prevalence-weighted baseline (see Baseline weighting methodology, above), sorted by co-recommended rate. Co-recommended rates span 16.5-39.2%: every platform is a statistically competitive choice in at least a sixth of scenarios. Maximising each of coverage, throughput, sensitivity, and pQTL in isolation, restricted to plasma (the dominant real matrix, so this isolates each dimension's effect from sample-type interactions entirely), still shifts the recommendation and separates cleanly: coverage to SomaSeq (25.0%), sensitivity to Alamar NULISA (33.3%), and pQTL to Seer Proteograph XT (100.0%, the most decisive single-dimension result in the simulation). Throughput is a genuine tie between Olink Explore HT and Nomic nELISA (20.8% each): both score 5/5 on throughput itself, and in plasma specifically neither platform's broader profile is strong enough to separate from the other, unlike in the full six-matrix sweep (used only as a cross-check, not reported in the figure) where Nomic's advantage in other, non-plasma-relevant dimensions gave it a narrow edge.
Seer Proteograph XT leads on both statistics (39.2% co-recommended, 12.3% outright), the clearest single result in the baseline: strong scores across multiple dimensions (coverage and pQTL both 5/5) translate into both frequent contention and, more often than any other platform, an uncontested lead. SomaSeq and Olink Explore HT are close behind on co-recommended rate (33.4%, 33.2%) but diverge sharply on outright win rate (9.4% vs. 3.1%): SomaSeq's coverage and precision scores (5/5 each) more often open a clean lead, while Olink's precision score (2/5) - independently evidenced by two published CV comparisons, see above - caps how often it separates from a crowded field even where it is competitive. Nomic nELISA (28.3% co-recommended, 6.7% outright) benefits from cost, throughput, and absolute-quantification strength; Alamar NULISA (17.4%, 4.7%) is concentrated in CNS-focus, CSF, and ultra-sensitivity scenarios, and is the only platform scoring exactly 1/5 on any of the seven user-adjustable dimensions (proteome coverage), so the only one the soft priority-miss penalty can affect via a slider dimension under the current, evidence-based score matrix. Biognosys TrueDiscovery has the lowest rate on both statistics (16.5%, 1.0%). Its strength is tissue: in a tissue-only slice it still has the highest outright-win rate (17.8%), though SomaSeq (13.0%) and Olink are much closer now that their tissue scores are corrected. Tissue is only about 6% of real sessions, so that strength barely moves its overall share. An earlier version of the engine, before the v3.0 multi-matrix-validation weight correction and the validation-based priority-miss extension described above, let an over-large tissue weight and uniform sample-type weighting inflate this narrow tissue strength into an apparent overall lead; those corrections removed that inflation.
Secondary check: plasma, discovery goal, large/extra-large study
As a secondary check, we evaluated the most prevalent real-world user profile: plasma matrix, biomarker discovery goal, large or extra-large study size (4,374 scenario-weight combinations, specialised toggles off), sorted by co-recommended rate. This narrow slice shows how recommendations concentrate when several study parameters are fixed to a single common context.
| Platform | Co-recommended | Outright win (slice) | Notes |
|---|---|---|---|
| Seer Proteograph XT | 45.0% | 17.6% | Maximal coverage (5/5) and pQTL (5/5) scores align with the discovery goal’s coverage boost |
| Olink Explore HT | 39.7% | 4.1% | Frequently tied here; rarely uncontested |
| Nomic nELISA | 36.7% | 9.3% | Cost and throughput boosts from large/xlarge size stack on its 5/5 scores |
| SomaSeq | 29.1% | 6.7% | Coverage- and precision-dominated discovery scenarios |
| Alamar NULISA | 8.1% | 0.1% | Coverage 1/5 is decisive against the discovery coverage boost |
| Biognosys TrueDiscovery | 0.0% | 0.0% | Multi-matrix validation advantage does not apply to plasma; coverage (4/5) trails the coverage-boosted leaders here |
Seer leads this slice on both statistics. Two of the discovery goal's three weight adjustments materially affect which platform leads within this narrow slice specifically: the coverage boost (+2, versus +1 for every other goal-adjusted dimension in the system) and the choice of specificity over sensitivity as the third boosted dimension - both are examined in Goal and context adjustment rationale, below. Reducing the coverage boost to +1 shifts the outright-win lead to Nomic nELISA (16.6% versus Seer's 12.9%); boosting sensitivity instead of specificity shifts it to SomaSeq (13.3% versus Seer's 11.9%). Neither adjustment moves Olink Explore HT's outright win rate by more than half a percentage point in this slice (4.1% to 4.3%, and 4.1% to 4.6%, respectively): its position here is set by its coverage score (4/5, one tier below Seer and SomaSeq) and its precision score (2/5, not offset by any adjustment active in this slice), not by either of these two adjustable parameters. Biognosys TrueDiscovery scores 0% on both statistics here: its multi-matrix validation advantage provides no benefit in plasma, and its coverage score (4/5, one tier below Seer and SomaSeq) keeps it out of the top group in a slice where the discovery goal boosts coverage. This is consistent with its prevalence-weighted baseline outright-win rate (1.0%, the lowest of the six), where its tissue-matrix advantage is real (see above) but tissue is a small enough share of real usage that it barely moves the overall figure. This narrow-slice concentration is expected: with the discovery goal fixed for all 4,374 scenarios here, any discovery-specific adjustment is active in every one of them, rather than the roughly one-sixth of baseline-grid scenarios where it applies.
Real-world validation: 65 anonymized survey respondents
Each of the 65 Help Me Choose questionnaires completed and downloaded by real users (April-June 2026) was rescored against the current engine using that respondent's own stated goal, sample type, size, priority weights, and toggles - 1 excluded as an internal test submission, 2 further excluded where the stated requirements hard-filtered every platform, leaving n=63, sorted by co-recommended rate.
| Platform | Co-recommended | Outright win | Notes |
|---|---|---|---|
| Seer Proteograph XT | 57.1% | 15.9% | Highest co-recommended rate of any platform in this dataset |
| SomaSeq | 44.4% | 0.0% | Narrowly ahead of Olink Explore HT; zero outright wins at n=63 is a small-sample floor (baseline: 9.4%) |
| Olink Explore HT | 42.9% | 0.0% | Narrowly behind SomaSeq; zero outright wins at n=63 is a small-sample floor (baseline: 3.1%) |
| Nomic nELISA | 33.3% | 19.1% | Cost, throughput, and the sole absolute-quantification platform |
| Alamar NULISA | 19.1% | 0.0% | Zero outright wins at n=63 is a small-sample floor (baseline: 4.7%) |
| Biognosys TrueDiscovery | 20.6% | 0.0% | More often co-recommended after the cost correction; still zero outright wins at n=63, and tissue (its best context) is only ~6% of respondents |
The real-world spread is still wider than the prevalence-weighted baseline grid even though both now reflect real-world prevalence, because real respondents' characteristics are correlated (a tissue-focused study is more likely to also share a particular goal or toggle) in a way the baseline grid's independent-axis weighting cannot capture, and real respondents' own slider choices are not a uniform sweep across weight levels the way the grid's are. Co-recommended rates run higher across the board here than in the baseline grid for the same reason: real inputs cluster more tightly than an independent-axis model predicts, so more platforms land within reach of each other at once. At n=63, four platforms show zero observed outright wins; the baseline grid confirms all four can win outright, just rarely, so this is a small-sample floor effect rather than evidence any of them structurally cannot win.
How often is there an outright winner at all? A minority of the time, in every dataset above: no platform stands alone in 62.9% of baseline scenarios, 62.1% of the secondary-check slice, and 65.1% of scored real-world respondents. This is a property of the scenario population, not of any one platform, which is why it is stated here rather than as a row inside the platform tables. The full breakdown by tie-group size (how often a tie is 2-way versus larger) is in the downloadable reproducibility bundle (choose_tie_distribution.csv, Reproducibility tab) rather than reproduced here; in short, ties shrink quickly by size in the prevalence-weighted baseline grid (each larger tie-group size is markedly less common than the last) but skew toward 3- and 4-way in real usage, consistent with real study designs clustering around a few common profiles more tightly than an independent-axis prevalence model predicts.
Goal and context adjustment rationale
Every automatic weight adjustment (goal, study size, sample type, the CNS and longitudinal toggles) is applied on top of the user's own slider values, never in place of them, and each is disclosed here individually rather than folded into an undifferentiated "calibration" step. Two have a dated, documented rationale in the version history: the discovery goal's coverage boost (+1 to +2 at v2.0, "reflecting proteome breadth as a core discovery requirement") and the population goal's coverage boost (+1, added at v2.0). Every other goal and context adjustment has been part of the engine since v1.0 without an individually recorded rationale, though each maps to the goal's own stated definition: drug target identification boosts coverage and specificity, matching its definition ("broad coverage with mechanistic resolution"); validation boosts precision and specificity, matching its definition verbatim; population boosts cost and throughput, matching "tens of thousands of samples, cost-controlled." This is a lower evidentiary bar than the platform scores, which cite specific published sources, and we are stating that plainly rather than presenting expert judgement as literature-derived.
The discovery goal specifically boosts coverage (+2, the only double-magnitude adjustment in the system - documented above), throughput (+1, reflecting that discovery studies increasingly run at cohort scale), and specificity (+1). The third choice does not clear even the face-validity bar the others do: no record explains why specificity was boosted rather than sensitivity. A case exists for either - specificity, to avoid biomarker leads that fail to replicate; sensitivity, to avoid filtering out weak-but-real signals before downstream validation can assess them - and neither has, to our knowledge, been weighed against actual discovery-phase proteomics methodology. We tested its effect directly rather than leaving the question open: swapping to a sensitivity boost shifts the outright-win leader of the plasma/discovery/large-xlarge secondary check above from Seer Proteograph XT to SomaSeq; it does not meaningfully change Olink Explore HT's position in that slice or in the full baseline grid. We have not changed this weight - reasonable alternatives exist and neither is required by, nor improves, any specific platform's position - but the gap is disclosed rather than left undocumented, and we would welcome a literature-grounded case for either choice.
Absolute quantification requirement
In v2.4 the absolute-quantification requirement became a capability filter rather than a score threshold. In the absolute-quantification-required slice, only platforms whose standard product reports validated physical-concentration output remain eligible: among the scored products, Nomic nELISA alone (ELISA-calibrated pg/mL), so Nomic wins 100% of that slice. Olink Explore HT, SomaSeq, Seer Proteograph XT, Alamar NULISA and Biognosys TrueDiscovery are all excluded; several of these vendors offer absolute quantification through separate products not scored here (Olink Flex / Target 48, Alamar NULISAseq AQ, Biognosys TrueSignature). In observed survey usage, an absolute-quantification requirement was set by only ~11% of sessions (required 10.8%, preferred 50.8%, not needed 38.5%; n=65), so this filter affects a minority of users while correctly steering them to the one scored platform that meets the requirement. The graded quantification-type scores (Nomic 4; Seer and Biognosys 3; Olink, SomaSeq and NULISA 2) still differentiate platforms in "nice-to-have" scenarios where absolute quantification is preferred but not required.
Known limitations of this analysis
- The optional protein-target search (Help Me Choose Step 1), including target lists populated by the Reactome pathway selector described below, adds a further weight and a coverage hard filter that neither sweep below exercises: the downloadable per-session results file does not record which proteins, if any, a respondent searched for or selected, so real-world usage of this path cannot be reconstructed and it is excluded from the reproducibility bundle rather than assumed to carry no effect.
- The real-world sample (n=63 scored respondents, of 65 downloaded) is a convenience sample of people who completed the Help Me Choose questionnaire and downloaded their results between April and June 2026, not a random sample of proteomics researchers.
- Co-recommended rate and outright win rate answer different questions and are not interchangeable: co-recommended rate (the metric led above) asks how often a platform is tied with at least one other platform at the top; outright win rate asks how often it is the sole, uncontested best choice. Both are reported for every dataset; neither alone is the full picture. Co-recommended rates can sum to well over 100%, since every platform in a multi-way tie is counted for that scenario; outright win rates do not sum to 100% (the remainder is scenarios with no outright winner at all; see "How often is there an outright winner at all?" above and the tie-group-size breakdown in the reproducibility bundle).
- Goal and context weight magnitudes carry a lower evidentiary bar than the platform scores; see Goal and context adjustment rationale, above, for the full disclosure of which are individually documented and which are not.
- The prevalence-weighted baseline (see Baseline weighting methodology, above) depends on the 65-respondent sample being reasonably representative of true usage, a dependency the previous uniform methodology did not have. It also assumes goal, sample type, study size, and toggle state are statistically independent (weighted as a product of their individual observed rates); real respondents' characteristics are likely correlated, which is part of why the real-world dataset's spread still differs from the baseline grid's even though both now use real-world prevalence.
Platform Combination Engine (Help Me Combine)
The Help Me Combine tab answers a different question from Help Me Choose: many studies run more than one platform, so which two technologies best complement each other? It reuses the same Atlas dimension scores (each rated 1 to 5) as inputs, rather than creating new platform scores, and evaluates the fifteen pairs formed from all six coverage-bearing platforms. Seer Proteograph XT and Biognosys TrueDiscovery are modelled as two independent mass-spectrometry platforms, each scored against its own accession-level coverage list (Biognosys TrueDiscovery's list is the published Kirsher et al. 2025 depletion-workflow dataset described above); they were previously merged into a single "mass spectrometry" category before that list existed.
Each pair is scored on several user-weightable axes: coverage against a selectable denominator (ChEMBL drug targets, the SwissProt canonical proteome, the Atlas detectable-proteome union, or FDA-approved biomarkers), complementarity measured as the net-new targets the second platform adds beyond the better single platform (rather than a unique-fraction ratio, which would reward pairing a small panel with a large one for the wrong reason), combined measurement score, protein-class balance, and combined cost. Coverage is computed exactly for all four numeric denominators, not estimated: drug-target and FDA-biomarker coverage are exact set intersections against their external reference lists, and canonical-proteome/Atlas-union coverage are exact unions of each platform's own accession list. The user selects up to three priorities, for example coverage, target specificity, or throughput, which set the axis weights; no axis is privileged by default, so coverage leads only if it is chosen. Sample volume is opt-in, and the combined measurement score can be recomposed from its eight underlying measurement properties, including maturity properties such as throughput that are off by default. An advanced panel exposes every weight directly.
The companion manuscript reports 15,123 protein identifiers and Protein Coverage lists 15,128 catalogue entries; the difference is the five phospho-tau rows that NULISA lists under MAPT (P10636).
Because the pairs cluster closely, the engine reports a tiered recommendation rather than a single winner. Coverage percentages carry bootstrap confidence intervals, pairs whose intervals overlap are presented as co-leading rather than ranked, and robustness is summarised by how often a pair lands in the top three across many weightings. Coverage here is list membership against the chosen denominator, an upper bound on usable coverage, and combined measurement score and cost are the Atlas expert ratings. The full methodology, including the coverage computation and the axis-redundancy and robustness analyses, is described in a companion preprint.
A permutation sweep tests whether the framework structurally favours any platform. Across every combination of up to three priorities, evaluated over the four numeric denominators (924 scenarios), the recommendation is well balanced and, with six platforms in play, no pair reaches the leading-choice (Tier 1) threshold: the top pair, Nomic Omni + Seer Proteograph XT, wins 18.7% of scenarios and leads the field on top-3 share (48.5%), but every platform appears in the winning pair in 11% to 52% of scenarios (Nomic Omni 52%, Seer Proteograph XT 50%, Olink Explore HT 37%, SomaSeq 26%, Alamar NULISA 24%, Biognosys TrueDiscovery 11%). An unconstrained sweep of 160,000 random weightings was also tested during development and later withdrawn: two of the seven axes, sample volume and mass-spec tissue fit, are zero-weighted by default and reachable only when a user explicitly selects them, so uniform random draws gave them the same expected influence as coverage, a weighting the priority picker itself cannot produce; it is not part of this analysis or the reproducibility bundle. Platform participation is substance-driven, not bias: Seer Proteograph XT leads solo coverage on two of the four denominators (canonical proteome, Atlas union), and Nomic Omni, despite the second-lowest solo coverage of the six platforms on every denominator (Alamar NULISA is lowest), gains the most from pairing precisely because of that gap, on top of being the consortium's default anchor. Biognosys TrueDiscovery, mid-table on solo coverage across all four denominators, is the newest entrant and appears least in the priority sweep, even though it produces the single largest net-new contribution of any Nomic Omni pairing (140 net-new drug targets). The engine specification, reproduction script, and the sweep and bootstrap-CI outputs are downloadable from the Reproducibility tab.
Pathway Selector (Reactome)
Both Help Me Choose and Help Me Combine let users narrow to a specific biological pathway instead of, or alongside, manually listing target proteins. Pathway membership is sourced from Reactome (811 named human pathways with ten or more members present in the canonical proteome, retrieved via the Reactome ContentService); selecting one resolves to that pathway's gene list.
In Help Me Choose, the resolved list enters the exact same input as a manually typed gene list, carrying the same weight of 6 and the same coverage hard filter (a platform is excluded if it covers under 50% of three or more listed targets). In Help Me Combine, pathway coverage is shown only as a descriptive figure alongside each pair's existing composite score, with an optional toggle to re-order, not rescore, the ranked list by it; it is not one of the seven composite axes, not one of the eleven priorities, and not a coverage denominator, so neither engine's scoring changes.
Reactome annotation is uneven across biology: it is denser for extensively studied pathways, so high coverage of a sparsely annotated pathway is a weaker statement than the same figure for a well-annotated one, since sparse annotation may reflect incomplete curation rather than the biology being genuinely hard to measure.
This code path is not exercised by either engine's fairness sweep, for the same reason the existing protein-target search is not (see Known limitations, above): the downloadable results format does not record which genes, if any, a respondent searched for or selected. Routing pathway selection through that same existing input does not close this gap; it inherits it.
Limitations & Disclosures
This tool focuses on commercially available, high-plex, untargeted or semi-targeted proteomics platforms with published head-to-head comparison data in human samples. Targeted panels (e.g., ProcartaPlex, LEGENDplex), single-protein assays (ELISA, Simoa), and research-use-only or early-access platforms (e.g., Nautilus) are outside the current scope. Where a vendor offers multiple platforms, we evaluated the product best suited to large-scale discovery proteomics; other products within the same portfolio (e.g., Olink Reveal) were excluded to keep comparisons focused and avoid redundancy. Sample preparation and enrichment providers (e.g., PreOmics) were not evaluated as they supply upstream workflows rather than end-to-end proteomics platforms with their own quantification readout. The platform set will be expanded in future releases as new technologies reach commercial maturity and accumulate independent validation data.
Scores are relative to the current platform set, not absolute quality verdicts. A score of 3/5 on any dimension means that platform sits in the middle of the evaluated field on that metric, not that it performs poorly in absolute terms. All six platforms are actively used in peer-reviewed research and are appropriate for the right study design. Scores will be revised as new independent evidence emerges, and all changes are documented in the version history.
This tool provides general guidance based on published evidence. Platform performance varies substantially by study design, sample handling, freeze-thaw cycles, and specific protein targets. Always consult with platform vendors and review the primary literature for your specific use case.
Independence policy: Ignition Scientific does not consult for any proteomics platform vendor. This policy is maintained to preserve the independence and credibility of this tool's assessments.
Pricing ranges are approximate list-price estimates and may not reflect current negotiated rates, academic discounts, or volume pricing. Contact vendors directly for study-specific quotes.
Some studies cited in the Evidence Base were conducted by authors with disclosed relationships to platform vendors. See individual study publications for full conflict of interest disclosures. Study inclusion is based on methodological quality, not author affiliation.
Contact
For custom platform selection analysis, study design consulting, or questions about this tool:
Response within 2 business days. Custom analyses available for pharmaceutical, biotech, and academic groups.
Version History
November 4, 2026
No score, weight or recommendation changed. The Eldjarn et al. 2023 sample size now uses the paper's own UK Biobank figure, 54,265 (Sun et al. 2023 report 54,219 for the same cohort), which moves the any-platform participant total from 95,617 to 95,663. The site's data are now published as versioned files with JSON Schemas on a new Data page, under CC BY 4.0, and key findings are stored as structured claims with their source locations. The confidence intervals stored in the Help Me Combine specification now match those the site shows, which did not change; FDA-biomarker intervals were added to the published files; and the bootstrap notes now state the actual denominators, 1,449 drug targets and 216 FDA biomarkers. An MCP server, aptatlas-mcp, lets AI assistants and other tools query the same data and run the same recommendation engines as this site.
November 4, 2026
Corrections to the evidence base and headline numbers; no score, weight or recommendation changed. The Overview figure '170K+ participants' was wrong: it drew on parent-cohort sizes rather than the people in each comparison (for example 100,000 for PRECISE-SG100K, where 500 participants were profiled). It is replaced by ~6,600 participants measured on two or more platforms; the Evidence Base now gives per-study counts and the any-platform total. Sample sizes and eight citations in the Evidence Base were corrected against the papers. The Overview's largest-platform figure now counts distinct proteins (10,065 for Seer; the 10,698 protein groups Seer lists remain on its card with a note). Protein Coverage now labels what each count means: 14,340 distinct human proteins, listed under 15,128 catalogue entries. Its 'unique to one platform' and 'measured by all six' filters now work at protein level (5,722 and 118 proteins; previously 6,725 and 109 entries). Overview figures are computed from the data at build time. All 51 key findings in the Evidence Base were then checked against the full text of the cited papers; 44 were changed, including 14 with no support in the cited paper. The same claims were corrected in the platform details, Executive Summary and Help Me Choose: for example, NULISA's limit of detection is 10 aM for HIV p24; the published nELISA study used a 191-plex panel in cell culture samples and reports seven orders of magnitude, while the roughly 1,000-plex panel and ~$50 price are vendor figures; and Pietzner et al. 2026 detected 8,079 Seer protein groups, not ~8,511. Sissala et al. 2025 is now tagged as other mass spectrometry rather than Seer. The Methods build steps now state the actual sequence: Claude drafted the first scores, and the authors scored the platforms before opening that draft, then compared the two sets and adjudicated. Every score change, before and since release, was decided by Whelan and Smith-Byrne; Claude made edits only on their instruction. The prompt file generated in the same session as the draft also told the build to adjust scores so that platform totals fell between 28 and 33; the authors did not set this range or use platform totals when scoring or adjudicating. The Overview and Evidence Base now count independent head-to-head studies: the same samples on two or more high-plex platforms, not vendor-authored (criterion in Methods). Nicholas et al. 2026 was added as an independent head-to-head study, and Rea Reyes et al. 2025, Zeng et al. 2024 and Suhre et al. 2025 as foundational rows; vendor authorship is marked. The Overview has a new By the numbers strip and a simpler hero, its search now covers the full platform catalogue, and its Technology and Cost filters were removed. Across the site, the build-step circles and the Help Me Choose progress steps are clearer, Platform Comparison cards show the technology in place of a repeated score, All Platforms gains a Technology filter, a Columns control and a cost note column, and long detail views show when there is more to scroll. Every page now has its own address (/compare, /methods and so on), title and description, and its text is in the page before any script runs, so search engines and link previews can read it. Old links with # in them still work. On phones and tablets the current tab stays in view, detail panels fill the screen, and Help Me Choose starts each step at the top. Small grey and coloured text is darker and easier to read, and panels, sliders and options now work by keyboard and are named for screen readers.
October 10, 2026
Display: platforms with an evidence depth of 2/5 or lower are now flagged as lower confidence wherever their scores appear (Platform Comparison cards, detail panels and table; the Overview; and Help Me Choose and Help Me Combine results), with a note that their scores are provisional. The flag is computed from the evidence-depth score, so it follows any future score change. No score, weight or recommendation changed.
October 9, 2026
Text correction: removed statements that Alamar NULISA's p-tau217 assay is FDA-cleared. NULISA's p-tau217 assays are for research use only; the Overview, Help Me Choose, and NULISA platform profile now describe its attomolar sensitivity and well-validated blood p-tau217 measurement instead. Also corrected the Help Me Combine description of the FDA-approved biomarker set from 217 to 216 proteins, the distinct-protein count its coverage values have used since v4.2. No score, weight, or recommendation changed.
Evidence base and scoring are reviewed monthly. Platform additions and major score revisions are noted in the changelog.