Variant Pathogenicity Analysis
CADD, REVEL, AlphaMissense, ESM-2, EVE, SpliceAI with Bayesian ACMG aggregation
SETBP1 G870S
L2=10.75
Highest embedding disruption
DNMT3A R882H
LLR=-8.38
PP3_Strong
SETBP1 G870S
LLR=-9.80
PP3_Strong
Novel Variant
EZH2 V662A
0 PubMed hits · 0 / 20,739 GENIE samples with EZH2 on panel
Embedding Disruption
SETBP1 G870S 1
[1] R 2013
Recurrent SETBP1 mutations in atypical chronic myeloid leukemia. Nat Genet (2013)
2[2] H 2013
Somatic SETBP1 mutations in myeloid malignancies. Nat Genet (2013)
disrupts the SKI domain degron motif,
producing the largest embedding perturbation (L2=10.75,
cosine=0.409) among all patient variants.
G870 lies in the beta-TrCP1 degron that targets SETBP1 for ubiquitin-mediated
degradation; G870S abolishes beta-TrCP1 binding, and the stabilised protein
lowers PP2A activity (Piazza et al. 2013).
Pathogenicity Score Comparison
| Variant | CADD | REVEL | AlphaMissense 19 [19] J 2023 Accurate proteome-wide missense variant effect prediction with AlphaMissense. Science (2023) | ESM-2 LLR 18 [18] Z 2023 Evolutionary-scale prediction of atomic-level protein structure with a language model. Science (2023) | EVE | SpliceAI | ACMG Points | Classification |
|---|---|---|---|---|---|---|---|---|
| DNMT3A R882H | 28.5 | 0.742 | 0.9953 | -8.383 | 0.6197 (Uncertain) | 0.01 | 20 | Pathogenic |
| IDH2 R140Q | 32.0 | 0.891 | 0.9872 | -1.478 | 0.8863 (Pathogenic) | 0.00 | 21 | Pathogenic |
| SETBP1 G870S | 28.7 | 0.716 | 0.9962 | -9.804 | 0.7460 (Uncertain) | 0.00 | 21 | Pathogenic |
| PTPN11 E76Q | 28.6 | 0.733 | 0.9972 | -1.865 | 0.3068 (Uncertain) | no score | 15 | Pathogenic |
| EZH2 V662A | 28.1 | 0.781 | 0.9984 | -2.966 | 0.7825 (Uncertain) | no score | 9 | Likely Pathogenic |
Values read from acmg_aggregation.json and eve_scores.json as regenerated
on 2026-11-03; CADD values are from Ensembl VEP, and REVEL from the
Ensembl VEP plugin except SETBP1 G870S, from myvariant.info (legacy pipeline).
EVE returns Uncertain for 4 of the five variants
(DNMT3A, SETBP1, PTPN11, EZH2) and Pathogenic only for
IDH2 R140Q; its calls
do not follow the gain- versus loss-of-function split. A SpliceAI entry of
"no score" means the allele was never pre-computed, not that the score is
low; BP7 is only applied where a delta score exists.
ACMG Evidence Aggregation
11 evidence sources were aggregated using the Bayesian point system
(Tavtigian et al. 2020) 26
DNMT3A R882H: 20 points (PS1 Strong, PS3 Strong, PM1 Strong, PP3 Strong, PP5 Strong; PP3_Strong on 5/7 concordant predictors).
IDH2 R140Q: 21 points (PS1 Strong, PS3 Strong, PM1 Strong, PM2 Supporting, PP3 Strong, PP5 Strong; PP3_Strong on 6/7 concordant predictors).
SETBP1 G870S: 21 points (PS1 Strong, PS3 Strong, PM1 Strong, PM2 Supporting, PP3 Strong, PP5 Strong; PP3_Strong on 6/7 concordant predictors).
PTPN11 E76Q: 15 points (PS1 Strong, PS3 Strong, PM1 Moderate, PM2 Supporting, PP3 Strong; PP3_Strong on 6/7 concordant predictors).
EZH2 V662A: 9 points (PS3 Supporting, PM1 Moderate, PM2 Supporting, PM5 Supporting, PP3 Strong; PP3_Strong on 6/7 concordant predictors).
IDH2 R140Q and SETBP1 G870S score highest, tied at 21 points. SETBP1 G870S disrupts the SKI domain degron motif 1
Four variants reach the Pathogenic threshold (≥10 points) and EZH2 V662A is Likely Pathogenic (9 points; research classification, ClinVar: VUS), on the aggregation regenerated 2026-11-03.
[26] SV 2020
Fitting a naturally scaled point system to the ACMG/AMP variant classification guidelines. Hum Mutat (2020)
for ACMG/AMP variant classification:
DNMT3A R882H: 20 points (PS1 Strong, PS3 Strong, PM1 Strong, PP3 Strong, PP5 Strong; PP3_Strong on 5/7 concordant predictors).
IDH2 R140Q: 21 points (PS1 Strong, PS3 Strong, PM1 Strong, PM2 Supporting, PP3 Strong, PP5 Strong; PP3_Strong on 6/7 concordant predictors).
SETBP1 G870S: 21 points (PS1 Strong, PS3 Strong, PM1 Strong, PM2 Supporting, PP3 Strong, PP5 Strong; PP3_Strong on 6/7 concordant predictors).
PTPN11 E76Q: 15 points (PS1 Strong, PS3 Strong, PM1 Moderate, PM2 Supporting, PP3 Strong; PP3_Strong on 6/7 concordant predictors).
EZH2 V662A: 9 points (PS3 Supporting, PM1 Moderate, PM2 Supporting, PM5 Supporting, PP3 Strong; PP3_Strong on 6/7 concordant predictors).
IDH2 R140Q and SETBP1 G870S score highest, tied at 21 points. SETBP1 G870S disrupts the SKI domain degron motif 1
[1] R 2013
Recurrent SETBP1 mutations in atypical chronic myeloid leukemia. Nat Genet (2013)
2[2] H 2013
Somatic SETBP1 mutations in myeloid malignancies. Nat Genet (2013)
through a
well-characterised gain-of-function mechanism. EZH2 V662A is a novel
unreported variant in the catalytic SET domain.
Four variants reach the Pathogenic threshold (≥10 points) and EZH2 V662A is Likely Pathogenic (9 points; research classification, ClinVar: VUS), on the aggregation regenerated 2026-11-03.
EZH2 V662A: Novel Unreported Variant
EZH2 V662A returns zero results in PubMed indexed literature. This variant
has never been described in a published case report. Chase et al.
(Leukemia, 2020, PMID 32322039)
10
Ball et al. (2023) 11
Clinical implication: Tazemetostat (EZH2 inhibitor, formerly FDA-approved in follicular lymphoma; Ipsen withdrew it from all its markets on March 9, 2026 after hematologic second cancers, mostly MDS and AML) is CONTRAINDICATED. Tazemetostat inhibits EZH2 catalytic activity, which V662A is predicted to have lost. Administering it would have no therapeutic benefit and should be flagged in any treatment discussion.
[10] A 2020
Mutational mechanisms of EZH2 inactivation in myeloid neoplasms. Leukemia (2020)
and the same group's earlier in vitro work
showed loss or greatly reduced H3K27 methylation for the SET/CXC missense
mutants they tested (C576W, R690C, Y731D). V662 sits in the catalytic SET
domain but was not tested, so loss of function for V662A is inferred, not
measured. EZH2 loss-of-function in myeloid neoplasms was first
characterized by Ernst et al. (2010)
9[9] T 2010
Inactivating mutations of the histone methyltransferase gene EZH2 in myeloid disorders. Nat Genet (2010)
, establishing
EZH2 as a tumour suppressor in this context.
Ball et al. (2023) 11
[11] S 2023
Clinical characteristics and outcomes of EZH2-mutant myelodysplastic syndrome: a large single institution analysis of 1774 patients. Leuk Res (2023)
reported that EZH2-mutant MDS patients
have inferior overall survival and higher rates of AML transformation,
underscoring the prognostic significance of this variant class.
Clinical implication: Tazemetostat (EZH2 inhibitor, formerly FDA-approved in follicular lymphoma; Ipsen withdrew it from all its markets on March 9, 2026 after hematologic second cancers, mostly MDS and AML) is CONTRAINDICATED. Tazemetostat inhibits EZH2 catalytic activity, which V662A is predicted to have lost. Administering it would have no therapeutic benefit and should be flagged in any treatment discussion.
Experimental Functional Validation (DMS)
PTPN11 E76Q: experimentally confirmed gain-of-function.
Deep mutational scanning of full-length SHP2 (12,054 variants)
29
All 19 possible amino acid substitutions at position E76 are gain-of-function (enrichment range 0.165–0.469). E76 is a critical N-SH2 autoinhibitory contact: any change disrupts the closed conformation and constitutively activates the phosphatase. The most common clinical variant, E76K, shows 188-fold increased catalytic activity (kcat/KM) over wildtype.
This is the difference between "predicted pathogenic" and "experimentally measured pathogenic." Together with DNMT3A R882H DMS data (Garcia et al. 2025, a bioRxiv preprint), 2 of 5 driver mutations have deep mutational scanning data. In the ACMG aggregation, PS3 is Strong for 4 of 5 variants (DNMT3A, IDH2, SETBP1, PTPN11) and Supporting for EZH2.
[29] Z 2025
Deep mutational scanning of the multi-domain phosphatase SHP2 reveals mechanisms of regulation and pathogenicity. Nat Commun (2025)
provides direct experimental evidence, not a computational prediction.
E76Q scores enrichment 0.329 (99th percentile,
z=3.70), classifying it as
gain-of-function with PS3_Strong ACMG evidence, a Strong-level
criterion worth 4 points on the Bayesian scale.
All 19 possible amino acid substitutions at position E76 are gain-of-function (enrichment range 0.165–0.469). E76 is a critical N-SH2 autoinhibitory contact: any change disrupts the closed conformation and constitutively activates the phosphatase. The most common clinical variant, E76K, shows 188-fold increased catalytic activity (kcat/KM) over wildtype.
This is the difference between "predicted pathogenic" and "experimentally measured pathogenic." Together with DNMT3A R882H DMS data (Garcia et al. 2025, a bioRxiv preprint), 2 of 5 driver mutations have deep mutational scanning data. In the ACMG aggregation, PS3 is Strong for 4 of 5 variants (DNMT3A, IDH2, SETBP1, PTPN11) and Supporting for EZH2.
| Variant | Enrichment | Z-score | Classification | Clinical Database |
|---|---|---|---|---|
| E76L | 0.469 | - | GoF | - |
| E76S | 0.436 | - | GoF | - |
| E76R | 0.436 | - | GoF | - |
| E76M | 0.399 | - | GoF | COSMIC |
| E76Q | 0.329 | 3.70 | GoF | COSMIC+TCGA+ClinVar |
| E76D | 0.165 | - | GoF | ClinVar |
| WT | 0.000 | - | Neutral | - |
| E76stop | -0.017 | - | Neutral | - |
Jiang et al. Nat Commun 2025 (PMID 40595497)
29
[29] Z 2025
Deep mutational scanning of the multi-domain phosphatase SHP2 reveals mechanisms of regulation and pathogenicity. Nat Commun (2025)
.
Deep mutational scanning of full-length SHP2 in a yeast (S. cerevisiae) growth-rescue assay with Src co-expression.
Enrichment scores validated against biochemical kcat/KM (Pearson r=0.718).
Site Conservation
Values are ESM-2 predicted per-position entropy (bits) from the deep run: low
entropy means the model is near-certain of the wildtype residue. They are model
predictions, not measured from an ortholog alignment.
SETBP1 position 870: entropy 0.0092 bits (very low). The model all but fixes the glycine at 870, consistent with its place in the SKI domain degron motif.
DNMT3A position 882: entropy 0.0279 bits. Low, though not the lowest in the scanned window 832-912 (position 834: 0.0028 bits); R882 is a known mutational hotspot in myeloid malignancies.
EZH2 position 662: entropy 2.4850 bits. PTPN11 position 76: entropy 2.6186 bits. IDH2 position 140: entropy 3.8056 bits. At these positions the model accepts several residues, so their pathogenicity is not captured by sequence constraint alone.
SETBP1 position 870: entropy 0.0092 bits (very low). The model all but fixes the glycine at 870, consistent with its place in the SKI domain degron motif.
DNMT3A position 882: entropy 0.0279 bits. Low, though not the lowest in the scanned window 832-912 (position 834: 0.0028 bits); R882 is a known mutational hotspot in myeloid malignancies.
EZH2 position 662: entropy 2.4850 bits. PTPN11 position 76: entropy 2.6186 bits. IDH2 position 140: entropy 3.8056 bits. At these positions the model accepts several residues, so their pathogenicity is not captured by sequence constraint alone.
Methods
Model: Scoring was performed using
facebook/esm2_t33_650M_UR50D 18
Hardware: NVIDIA GeForce RTX 4060 (8 GB VRAM); about 2.4 s for the whole basic scoring run on GPU, 58.4 s for the deep run.
Technique: Masked marginal log-likelihood ratio scoring. For each variant position, the wildtype and mutant residues are masked, and the model computes the probability of observing each amino acid given the sequence context. The LLR is the log-ratio of mutant vs wildtype probability. More negative LLR indicates the mutation is more unexpected by the protein language model 18
PP3 classification: LLR cut-offs set in this project's scoring script, based on the ESM1b benchmarks of Brandes et al. 2023 20
[18] Z 2023
Evolutionary-scale prediction of atomic-level protein structure with a language model. Science (2023)
(650M parameters,
33 transformer layers, trained on UniRef50) with the masked marginal
log-likelihood ratio approach 20[20] N 2023
Genome-wide prediction of disease variant effects with a deep protein language model. Nat Genet (2023)
.
Hardware: NVIDIA GeForce RTX 4060 (8 GB VRAM); about 2.4 s for the whole basic scoring run on GPU, 58.4 s for the deep run.
Technique: Masked marginal log-likelihood ratio scoring. For each variant position, the wildtype and mutant residues are masked, and the model computes the probability of observing each amino acid given the sequence context. The LLR is the log-ratio of mutant vs wildtype probability. More negative LLR indicates the mutation is more unexpected by the protein language model 18
[18] Z 2023
Evolutionary-scale prediction of atomic-level protein structure with a language model. Science (2023)
.
PP3 classification: LLR cut-offs set in this project's scoring script, based on the ESM1b benchmarks of Brandes et al. 2023 20
[20] N 2023
Genome-wide prediction of disease variant effects with a deep protein language model. Nat Genet (2023)
:
PP3_Strong (LLR < -7.0),
PP3_Moderate (-7.0 to -3.5), PP3_Supporting (-3.5 to -1.5).
References
- Piazza R et al. Recurrent SETBP1 mutations in atypical chronic myeloid leukemia. Nat Genet (2013). PubMed
- Makishima H et al. Somatic SETBP1 mutations in myeloid malignancies. Nat Genet (2013). PubMed
- Lin Z et al. Evolutionary-scale prediction of atomic-level protein structure with a language model. Science (2023). DOI
- Brandes N et al. Genome-wide prediction of disease variant effects with a deep protein language model. Nat Genet (2023). DOI
- Cheng J et al. Accurate proteome-wide missense variant effect prediction with AlphaMissense. Science (2023). DOI
- Tavtigian SV et al. Fitting a naturally scaled point system to the ACMG/AMP variant classification guidelines. Hum Mutat (2020). DOI
- Chase A et al. Mutational mechanisms of EZH2 inactivation in myeloid neoplasms. Leukemia (2020). DOI
- Ernst T et al. Inactivating mutations of the histone methyltransferase gene EZH2 in myeloid disorders. Nat Genet (2010). PubMed
- Ball S et al. Clinical characteristics and outcomes of EZH2-mutant myelodysplastic syndrome. Leuk Res (2023). DOI
- Jiang Z et al. Deep mutational scanning of the multi-domain phosphatase SHP2. Nat Commun (2025). PubMed