Are O4 and O7 independent lines of evidence?

SummaryThe short answer

O4 and O7 are not fully independent, and the dependency is larger and more structural than the shared-cohort argument alone suggests. There are two distinct channels of double counting, and they call for different remedies.

1. A shared-cohort channel. Just under half of the mutation observations that make a CancerHotspots residue statistically significant come from MSK‑IMPACT, which is also a GENIE contributing centre. For a typical hotspot, roughly half the observations behind O7 are the same patient observations an analyst then counts in GENIE for O4.
2. A definitional channel. Every route to O7 through cancerhotspots.org in SVIG-UK v1.1 is gated on the count of the same amino-acid change — which is precisely what O4 measures. This holds irrespective of cohort overlap, and it is the finding we would put before the group first.
3. The coupling is measured, not assumed. In GENIE v20, counting one entry per patient as the guideline requires, 197 of the 198 amino-acid changes reaching O7_strong (99.5%) also reach O4_strong — and all 197 still do with every MSK‑IMPACT patient removed. Excluding the shared cohort does not prevent a single +8, so the double counting is definitional rather than a cohort artefact, and only a cap removes it.
4. Capping Cancer Hotspots at supporting across the board is not the right fix. It would discard genuine positional evidence at the ~18% of hotspot residues that are mutated across many different substitutions — KRAS p.G12, NRAS p.Q61, TP53 p.R273.
5. The conditional cap proposed by KRB is the right shape, it can be made operational, and its impact is small and targeted. It would leave the score of roughly 85% of hotspot changes untouched and drop only ten from +8 to +4, four of which are already covered by O1 and therefore unaffected in practice.

Section 1The question, and why it is not obvious

O4 and O7 are intended to capture different things. O4 measures recurrence and enrichment of a specific variant in a large somatic dataset. O7 asks whether the residue is a statistically significant mutational hotspot or lies in a critical functional region. On that reading they are separable lines of evidence, and Supplementary Figure 1 of SVIG-UK v1.1 permits them to be combined without restriction — up to +8 points.

The concern is that both are frequently satisfied by the same underlying patient observations. Cancer Hotspots is not a simple case-counting database; it applies a statistical model that accounts for mutability at a given amino acid and for gene- and position-specific mutation rates.^1,2 But if the residue's significance is driven largely by the same cases that are counted in GENIE, then awarding +4 for O4 and +4 for O7 gives +8 points from what is, biologically, partly one signal of positive selection.

The guideline is alert to this. Its O7 section already ends with an instruction to exclude the overlapping cohorts:

“MSK and the TCGA should be excluded from this count to enable the use of Cancer Hotspots (O7) without double counting evidence.”

SVIG-UK v1.1, main text, O7

The intent is exactly the one under discussion. But the sentence sits at the end of a paragraph about in silico protein modelling and DECIPHER plots, so “this count” has no unambiguous antecedent; and read as an instruction to exclude MSK and TCGA cases from the O4 count it is not practicable, since TCGA is not in GENIE and the GENIE portal offers no per-centre view aligned to the CancerHotspots cohort. The mitigation exists on paper but cannot be executed at the bench.

This report sets out to establish, from the data, how large the overlap actually is, whether it matters, and what a workable remedy would look like.

Section 2What was analysed

Three resources were used, all at the level of the amino-acid change, which is the level at which SVIG-UK is written.

Cancer Hotspots

Both published releases were merged: Chang et al. 2016¹ (459 residues) and Chang et al. 2018² (1,110 single-residue positions across 24,592 tumours, plus 55 in-frame indel regions). For each hotspot residue the resource publishes the total mutation count and the count for each individual amino-acid change — for example NRAS p.Q61 carries 422 mutations distributed as Q61R:204 | Q61K:142 | Q61L:46 | Q61H:27 | Q61P:2 | Q61*:1.

Critically, the 2018 release also reports, for every residue, how that total splits between MSK‑IMPACT and retrospective public cohorts (largely TCGA). This split is what makes the O4/O7 overlap directly measurable rather than a matter of argument. It was verified that the two components reconstitute the published total for all 3,004 rows.

The primary analysis set is the 2018 single-residue table restricted to protein-coding positions: 1,024 residues and 2,918 distinct amino-acid changes. Splice hotspots (86 positions) are excluded because O7 does not apply to them, and within the analysis set a further 454 — 247 nonsense or stop-loss changes and 207 synonymous ones — are marked ineligible for O7, which is reserved for missense and small in-frame variants. That leaves 1,649 changes that score O7 at some strength from cancerhotspots.org.

COSMIC Cancer Mutation Census

GENIE was not available when this analysis was first written, so Sections 6 and 7 use COSMIC in its place; Section 8 repeats them on GENIE itself. SVIG-UK explicitly permits COSMIC as an alternative resource for O4, and the Cancer Mutation Census (v104) publishes a per-substitution sample count drawn from a cohort that is largely not the CancerHotspots cohort. It is used here to measure how often O4 and O7 fire on the same variant — not to make classification calls on individual variants.

AACR Project GENIE

Release 20.0-public, supplied by KRB: 289,869 samples from 242,866 patients across twenty contributing centres, of which MSK‑IMPACT is the largest (97,019 patients). O4 is counted as distinct patients carrying the change. Each sample is assigned a lineage — myeloid, lymphoid or solid — from the OncoTree tissue of its tumour type.

SVIG-UK Canonical Variants List

The O1 list (Supplementary Table 3) was used to establish which of the affected variants are already covered by standalone evidence, and therefore unaffected by any change to O4/O7 scoring.

Section 3How much of Cancer Hotspots is MSK?

Across the 1,024 protein-coding hotspot residues, 25,814 mutation observations underpin the hotspot calls. Of these, 12,087 (46.8%) come from MSK‑IMPACT and 13,727 (53.2%) from retrospective public cohorts.

Two of the most frequently assessed lung variants sit at the extreme. EGFR p.T790M rests on 60 mutations at the residue, of which 58 (96.7%) are MSK; EGFR p.L858R on 144, of which 114 (79.2%) are MSK. Conversely, the retrospective component is not in GENIE, so roughly half of the hotspot evidence is genuinely independent of a GENIE case count. The degree of overlap is measurable per residue, not constant.

Section 4O7’s thresholds are variant-level, not positional

This is the finding that does not depend on any cohort-overlap argument, and it is visible on the face of Supplementary Table 1. Read the cancerhotspots.org tiers as an analyst applies them:

Every tier is gated on the recurrence count of the exact same amino-acid change. The supporting tier is defined on nothing else. The positional count only ever arbitrates between strong and moderate — it never, on its own, admits O7.

O4’s missense thresholds are > 10 entries for strong and 5–10 for moderate. So the ≥ 10 same-change requirement that lifts O7 above supporting sits almost exactly on O4’s strong threshold. A variant scoring O7_strong from cancerhotspots.org has, by construction, satisfied a recurrence test nearly identical to the one that gave it O4_strong — on a patient set that is around 47% shared.

Conceptually, O4 and O7 can be independent. As currently specified, the cancerhotspots.org route to O7 does not measure the positional signal in isolation.

Section 5How positional are hotspots in practice?

For O7 to carry genuinely positional information, a residue needs to be mutated to several different amino acids. It frequently is not.

The more informative measure is the fraction of a residue’s mutations accounted for by its commonest substitution. The median is 0.69 and the mean 0.70. Classifying residues on that fraction:

Two residues in five are ≥ 80% dominated by a single change; for those, O7 adds little to O4. But the remaining 18.5% are the opposite case, and they are the reason not to cap Cancer Hotspots across the board. KRAS p.G12 carries ten distinct substitutions over 2,175 mutations, the commonest (G12D) accounting for only 34.8%. NRAS p.Q61 carries six over 422 (top 48.3%); TP53 p.R273 six over 609 (top 41.2%). At these residues the hotspot designation survives comfortably without the variant under assessment, and O7 is telling the analyst something O4 is not.

Section 6Testing the coupling against a second database

The argument so far is structural. It can be tested. COSMIC CMC provides an O4-style recurrence count drawn from a largely different cohort; 2,570 of the 2,918 hotspot changes (88.1%) match a COSMIC substitution.

The recurrence signal that drives O7 and the recurrence signal that would drive O4 are strongly correlated even across two largely different cohorts. This is the point worth putting to the group: removing the MSK overlap entirely would not make these two codes independent, because they are measuring the same underlying phenomenon — positive selection on a specific residue change.

Applying O4’s Supplementary Table 1 missense thresholds to the COSMIC counts gives the practical consequence:

195 of the 198 changes reaching O7_strong (98.5%) also reach O4_strong, as do 192 of 203 O7_moderate changes (94.6%). At the top of the scale the two codes fire together almost deterministically. Because the guideline warns that COSMIC thresholds should be higher than GENIE’s without specifying them, the analysis was repeated with a conservative scheme (≥ 51 strong, 20–50 moderate); the coupling holds, with 185 of 198 (93.4%) still reaching O4_strong.

Section 7What a cap would actually cost

A cap is only worth adopting if its effect is proportionate. Modelled over all 2,918 amino-acid changes:

The cap is surgical, not disruptive. It leaves the score of roughly 85% of changes untouched, and only ten amino-acid changes lose the full four points. Those ten, with their status on the Canonical Variants List — O1 is standalone evidence, so a variant on that list is classified Oncogenic whatever O4 and O7 contribute:

Four of the ten are already covered by O1, where the cap is moot. That leaves six variants materially affected, each of which the group can review individually — and each of which is precisely the case the proposal is aimed at: a hotspot defined by a single amino-acid change, scored twice.

Section 8Confirmed against GENIE

Sections 6 and 7 used COSMIC because GENIE was not available when they were written. KRB has since supplied AACR Project GENIE release 20.0-public, so the O4 side has been re-run on the database SVIG-UK actually names — with every count de-duplicated to one entry per patient, as the guideline requires. Every conclusion holds, and the one question COSMIC could not answer now has an answer.

“The GENIE database contains multiple samples from a single patient … and therefore care should be taken to ensure that multiple samples from the same patient are only counted as a single entry.”

SVIG-UK v1.1, Supplementary Table 1, O4 guidance

De-duplication

Counting patients rather than samples reduces the hotspot-change observations from 298,087 to 273,014 (−8.4%) and lowers the count of 1,701 changes, but moves only 24 O7-scoring changes down an O4 tier, because O4’s thresholds are low relative to hotspot counts. The inflation is concentrated in haematology: myeloid patients contribute 1.98 samples each against 1.12 for solid tumours, and 3.1 at MSK, reflecting serial sampling. JAK2 p.V617F is 3,393 samples but 1,947 patients. For a rarer haematological variant near a threshold, counting samples instead of patients is the difference between tiers.

The leave-MSK-out recount

GENIE records the contributing centre for every sample, so O4 can be counted with MSK‑IMPACT removed. Once it is, O4 and O7 rest on disjoint patients: GENIE’s nineteen other centres on one side, and on the other CancerHotspots, whose non-MSK half is TCGA and other public cohorts that GENIE does not contain.

The correlations make the same point from the other direction. Per residue, the CancerHotspots count from its retrospective cohorts and the GENIE count from its non-MSK centres — two sets of patients with nobody in common — still correlate at ρ = 0.68, against 0.81 for the MSK counts on both sides, which are the same patients.

The coupling and the cost of the cap, on GENIE

The one O7_strong change not reaching O4_strong, GTF2I p.L424H, is in a gene GENIE panels do not sequence. The cap reaches more changes on GENIE than on COSMIC because GENIE is larger: more O7_supporting changes clear O4_strong, and each loses a single point (+5 to +4). The costly case does not move: the same ten changes drop from +8 to +4, and the same four are protected by O1.

Haematological representation in GENIE

Haematological neoplasms are about a tenth of GENIE, and they are disproportionately MSK: MSK‑IMPACT contributes 40% of GENIE’s patients but 74% of its lymphoid patients. Excluding MSK therefore costs haematology far more than solid tumours — lymphoid neoplasms fall from 5.4% of GENIE to 2.4%. (3,146 patients have samples in more than one lineage, so the rows sum to slightly more than 100%.)

The guideline lets O4 counting be limited to “on-target” tumour types and, for AML, names the options: AML only, myeloid disorders, or all haematopoietic disorders. CancerHotspots is pan-cancer and solid-dominated, so a haematological on-target count is where O4 and O7 stop drawing on the same patients. Counted on haematological patients only, 83 of the 197 O7_strong changes still reach O4_strong; on myeloid patients only, 47. That is the on-target rule working as intended — most CancerHotspots O7_strong changes are solid-tumour drivers.

The haematological drivers that CancerHotspots lacks entirely (Section 11) are well represented in GENIE:

For these the two codes fall out of step in the opposite direction: O4 is available at strong on any on-target definition, while O7 via cancerhotspots.org is not available at all. The double-counting question does not arise — but the gap in Section 11 stands, and GENIE’s myeloid subset is the natural source for the supplementary haematological list proposed in Section 12. ASXL1 p.G646Wfs*12 is the exception to the lineage pattern, with more than half its patients recorded as solid tumours; clonal haematopoiesis in the sequenced sample is the likely explanation, and a reason for care whenever GENIE’s pan-cancer counts are used for a CHIP-associated gene.

Method: lineage is the OncoTree tissue of each sample’s tumour type (release 2025-10-03). GENIE annotates eight genes on a different isoform from CancerHotspots (EZH2 p.Y646 is p.Y641 in GENIE); these are aligned by a per-gene residue offset, adopted only where it recovers at least three missense hotspot changes. 2,655 of the 2,918 hotspot changes (91.0%) are observed in GENIE.

Section 9Proposed amendment

This is the amendment suggested by KRB, with the “demonstrably supported by multiple independent amino acid changes” clause turned into a test that can be applied at the bench from the same web page the analyst is already reading.

The leave-one-variant-out test

Subtract the count for the variant under assessment’s own amino-acid change from the residue total. The residue provides independent evidence if the residual is ≥ 10 mutations and at least one other amino-acid change at that residue is itself recurrent. The ≥ 10 threshold deliberately mirrors the count SVIG-UK already uses to lift O7 above supporting.

Two calibrations were modelled. The permissive test (another change with ≥ 2 mutations) leaves 752 of 1,649 O7-scoring changes uncapped (45.6%); the strict test (another change with ≥ 10 mutations, independently meeting the guideline’s own bar) leaves 585 (35.5%). The permissive test is recommended, because the strict test additionally penalises residues such as KIT p.D816 where the positional signal is real but the absolute counts are small.

Worked examples

BRAF p.V600E is the instructive borderline. The residue is 92.9% dominated by V600E, so it reads as mutation-specific — yet 64 mutations across five other changes remain after removing V600E, which is itself hotspot-level recurrence. The rule correctly lets it combine. EGFR p.L858R is the clean opposite: one change, 144 of 144, four-fifths of it MSK. There is no positional evidence there that is not L858R’s own recurrence.

Two further points for the group

The precedent for a cap already exists. Supplementary Figure 1, note 6 reads:

“Restriction: Sum of points awarding from combining O6 and O7 is limited to 4 pts (strong) to avoid double counting key attributes of critical domains / residues which are also often measured by in silico tools.”

SVIG-UK v1.1, Supplementary Figure 1, note 6

The proposed O4 + O7 cap is the same construction applied to a more strongly correlated pair. The O4 × O7 cell of that matrix carries no footnote at all.

The existing mitigation should be revised, not left standing alongside a cap. The “MSK and the TCGA should be excluded from this count” sentence quoted in Section 1 is ambiguous in its placement and impracticable as written. If a cap is adopted, that sentence should be replaced rather than duplicated.

Section 10Recommendation

1. Adopt the conditional cap, with the leave-one-variant-out test as the operational definition of independence, added as a footnote to the O4 × O7 cell of Supplementary Figure 1 in the form given above.
2. Do not cap Cancer Hotspots at supporting across the board. It would under-credit the ~18% of residues carrying real positional evidence and weaken classification of genuinely position-driven drivers such as KRAS p.G12. The conditional cap achieves the same protection while changing the score of only 15% of hotspot changes.
3. Revise the “MSK and the TCGA should be excluded from this count” sentence, which is ambiguous and not practicable as written.
4. State explicitly that O7 via cancerhotspots.org is, as written, a variant-level test, and that the positional argument requires evidence of multiple distinct changes at the residue.

Section 11Limitations and open points

GENIE counts are pan-cancer

Section 8 replaces the COSMIC stand-in with GENIE v20, and every conclusion of Sections 6 and 7 holds. The GENIE counts are pan-cancer. A laboratory applying the guideline’s “on-target” option will count fewer entries; the analysis scripts produce per-lineage counts for every change so that it can. GENIE panels also differ in gene content: three hotspot genes (GTF2I, HIST1H3C, HIST1H3H) are not sequenced by any of them, so a zero there means “not tested”, not “not seen”.

Version currency and choice

CancerHotspots v2 (2018, 24,592 tumours) predates current GENIE releases. The MSK fraction measured here is the MSK share within the hotspot resource, which is the quantity that determines whether the hotspot call and the GENIE count rest on the same patients — but the absolute overlap with a present-day GENIE query will differ, and Section 8 now measures it directly against GENIE v20. Separately, the two CancerHotspots releases differ substantially: 459 residues in v1 against 1,024 in v2, with only 252 shared. The website serves v2. Whichever is used should be version-stamped in the report, as the guideline already requires for GENIE.

Haematological coverage is a real gap

The guideline notes in general terms that cancerhotspots.org coverage is more limited for haematological malignancies. Quantified against named targets, JAK2 p.V617F, MPL p.W515, CALR exon 9, NPM1 exon 12, ASXL1 and SETBP1 are absent from both the SNV and indel tables of CancerHotspots v2. O7 is simply unavailable through this resource for several of the most important haematological drivers, which may be worth stating explicitly in the guidance rather than leaving to the analyst to discover. GENIE, by contrast, covers all of them well (Section 8).

Scope of the analysis set

Splice hotspots (86 residues) are excluded, as O7 does not apply to them. Nonsense and stop-loss changes (247) and synonymous changes (207) are marked ineligible for O7, which is reserved for missense and small in-frame variants (a nonsense change also cannot combine O7 with O2); their counts still contribute to the published residue totals, as they do on the website. The 55 in-frame indel regions are ranges rather than single residues, so the “distinct changes at a position” construction does not transfer, and they are excluded from the tier analysis. Protein-level matching to COSMIC failed for 11.9% of changes, largely because the two resources use different reference transcripts — MYD88 p.L265P appears as p.L273P in COSMIC, and GNAS p.R201 does not join at all.

A correction to earlier versions

Versions 1.0 and 1.1 of this report scored synonymous changes for O7. CancerHotspots lists them at hotspot residues — KRAS p.G60G has 19 mutations — and 87 of the 207 had been given O7 at supporting or moderate. They are now excluded, as nonsense and stop-loss changes already were. That changed the O7 tier counts and the denominators of the independence test (1,736 → 1,649 O7-scoring changes). It changed none of the 198 changes reaching O7_strong, none of the cap-impact figures, and none of the ten changes that drop from +8 to +4.

A textual discrepancy worth correcting

The main text permits O7 at strong for “> 50 occurrences at the same position”; Supplementary Table 1 says “≥ 50 entries”. These differ at exactly 50. This analysis uses the Supplementary Table wording, which is the operational table.

Section 12What could be done next, and what it would need

None of what follows is required to act on the recommendation in Section 10 — the case for the conditional cap rests on the evidence already presented. These are the analyses that would strengthen it, close a gap it leaves open, or extend the same reasoning to parts of the guideline this work did not reach.

Done: the GENIE leave-MSK-out recount

This section originally led with the analysis that would settle the question outright: repeating the O4 side on GENIE with MSK‑IMPACT removed. KRB supplied the data and it is now Section 8. It did settle it — no change reaching O7_strong loses O4_strong without MSK, so excluding the shared cohort is not a remedy and the cap is.

The rest of the list