bio-clinical-databases-tumor-mutational-burden

## Version Compatibility Reference examples tested with: cyvcf2 0.30+, VEP 111+ (or snpEff 5.2+), pandas 2.2+, numpy 1.26+, LOHHLA 1.0+ (McGranahan 2017), DASH 1.0+ (Pyke 2022). v4.1 (May 2024) gnomAD is current for germline subtraction. Friends of Cancer Research TMB harmonization framework (Vega 2021 *Ann Oncol*) and ESMO 2024 (Mosele *Ann Oncol*) define the operational thresholds. Before using code patterns, verify installed versions match. If versions differ: - Python: `pip show <package>` then `help(module.function)` to check signatures - CLI: `<tool> --version` If code throws ImportError, AttributeError, or TypeError, introspect the installed package and adapt the example to match the actual API rather than retrying. TMB calculation requires VCF with VEP / snpEff / Funcotator consequence annotations; the panel size used as denominator MUST match the assay's actual scored region, NOT the panel's total content. # Tumor Mutational Burden; Calculation, Harmonization, ICI Eligibility **'Calculate TMB from this somatic VCF and apply ICI eligibility cutoff'** -> Count nonsynonymous coding variants passing VAF/depth/germline filters; divide by assay scored region in Mb; apply assay-calibrated TMB-H cutoff; integrate with MSI / HLA-LOH / neoantigen quality. - Python: `cyvcf2.VCF()` + VEP/snpEff consequence parsing + panel-size normalization - CLI: `bcftools view` filtering + custom counting - HLA-LOH: LOHHLA (McGranahan 2017 *Cell*) or DASH (Pyke 2022 *Nat Commun*) - Neoantigen quality: pVAC-tools, NetMHCpan-4.1, Luksza 2017 fitness model ## Regulatory and Trial Landscape | Event | Year | Threshold | Notes | |-------|------|-----------|-------| | **KEYNOTE-158 + FDA pembrolizumab pan-tumor approval** | 2020 | TMB-H >= 10 mut/Mb | FoundationOne CDx companion diagnostic; 10 cohorts | | **Friends of Cancer Research TMB harmonization Phase I (Merino 2020)** | 2020 | -- | 11 panels vs WES truth; 3-fold panel-specific differences | | **Friends of Cancer Research Phase II (Vega 2021)** | 2021 | Calibration equations | 19 platforms; per-assay calibration to WES-aligned TMB-Mb | | **ESMO 2024 (Mosele *Ann Oncol*)** | 2024 | TMB-H >= 10/Mb retained (tumour-agnostic, ESCAT IB) | Tumour-type limits per McGrail 2021 | | **KEYNOTE-189 (NSCLC + pembrolizumab + chemo)** | 2018 | -- | TMB-H did NOT enrich for benefit with chemo backbone | | **POSEIDON / KEYNOTE-021 / KEYNOTE-407** | 2019-2022 | -- | TMB inconsistent with chemo backbones | | **B-F1RST + BFAST Cohort C (bTMB)** | 2022 | bTMB >= 16/Mb | BFAST Cohort C FAILED primary endpoint | ## Friends of Cancer Research Harmonization: Cross-Panel Calibration Merino 2020 *J Immunother Cancer*: in silico panel sampling from TCGA WES truth showed panel-specific TMB can differ 3-fold for identical samples. Vega 2021 *Ann Oncol* derived per-panel calibration equations to translate panel TMB to WES-aligned TMB-Mb. **Per-panel calibration to FoundationOne 10/Mb sensitivity:** | Panel | Scored region (Mb) | Equivalent threshold for FDA 10/Mb pan-tumor | Fails when | |-------|---------------------|----------------------------------------------|-----------| | **FoundationOne CDx** | 0.8 Mb scored (NOT 1.1 Mb total) | 10 mut/Mb (FDA reference; F1CDx companion) | Using 1.1 Mb panel total inflates TMB ~37%; pipeline excludes synonymous (F1CDx includes them) | | **MSK-IMPACT v3** | 0.98 Mb | ~10 (full Vega 2021 calibration recommended) | Tumor purity < 30%; non-paired-normal mode | | **MSK-IMPACT v4** | 1.22 Mb | ~10 | -- | | **TruSight Oncology 500** | ~1.3 Mb scored (from 1.94 Mb total) | **7.8 mut/Mb** | Pipeline uses 10/Mb instead of the TMB2-calibrated 7.8 (Ramos-Paradas 2021) | | **Oncomine Tumor Mutation Load** | 1.2 Mb | **8.4 mut/Mb** | Pipeline uses 10/Mb instead of the TMB2-calibrated 8.4 (Ramos-Paradas 2021) | | **Caris MI Tumor Seek** | ~1.2 Mb |; (verify Caris docs) | -- | | **Tempus xT v3** | 0.6 Mb | -- | Below 0.8 Mb minimum reliability threshold | | **Predicine ATLAS** | ~0.6 Mb | -- | Below 0.8 Mb minimum; high sampling variance | **TMB =/= TMB across vendors.** Manuscripts that compare TMB across panels without per-assay calibration are unreviewable. Use the Vega 2021 calibration equations or WES re-projection. ## Variant-Counting Subtleties These choices alter TMB by 5-20%: | Variable | Convention | Notes | |----------|-----------|-------| | **Synonymous variants** | **FoundationOne CDx INCLUDES synonymous** (rationale: reduces sampling noise); MSK-IMPACT and most academic pipelines exclude | The FDA companion diagnostic counts synonymous; frequent misconception | | **Indels** | FoundationOne includes; some assays exclude frameshift only | 5-15% TMB impact | | **Germline subtraction** | Paired-normal (gold standard); else gnomAD AF <=0.5% (sometimes 1%) for tumor-only | Population-stratified gnomAD AF for ancestry-diverse cohorts | | **VAF threshold** | FoundationOne >=5%; >=10% for tumor-only no UMI; down to 2% with paired-normal | Lower VAF risks contamination/artifacts | | **Hotspots** | COSMIC-confirmed driver hotspots typically EXCLUDED (not random) | Inflates TMB if included | | **Tumor purity** | FoundationOne >=20%; MSK-IMPACT >=30% | Below floor erodes VAF-based filtering | | **VEP version** | Pin to assay's annotation version | gnomAD v4 uses VEP 105 | ## Hypermutator Tiering | Class | Threshold | Common etiology | |-------|-----------|----------------| | **TMB-H (FDA pan-cancer)** | >= 10 mut/Mb | Variable; ICI eligible | | **Hypermutator (research)** | >= 100 mut/Mb | MMR-D, POLE-exo | | **Ultra-hypermutator** | >= 500 mut/Mb | POLE+MMR concurrent | MSI-H typically 30-50 mut/Mb; pure POLE-exo P286R 100-300 mut/Mb; POLE-exo + MMR-D exceeds 500. MSI-H and TMB-H overlap substantially (~83% of MSI-H are TMB-H) but only ~16% of TMB-H solid tumors are MSI-H (Chalmers 2017 *Genome Med* 9:34). ## The Tumor-Type-Specific Cutoff Debate **McGrail 2021** *Ann Oncol* is the most damning paper for the universal 10/Mb cutoff. TMB-H predicts ICI response in melanoma, NSCLC, bladder; but FAILS in breast, prostate, glioma. ORR in TMB-H melanoma/NSCLC/bladder was 39.8%; TMB-H breast/prostate/glioma was 15.3%. Mechanistic explanation: TMB only predicts when baseline CD8 T-cell infiltrate is present. **Sha 2020** *Cancer Discov*: TMB-H predicts ICI benefit in MSS subset but adds nothing on top of MSI-H (because MSI-H is uniformly hypermutator and uniformly responsive). **Samstein 2019** *Nat Genet* (MSK-IMPACT 1,662 ICI-treated): cancer-specific TMB cutoffs (top 20% within each tumor type) outperform universal 10/Mb. **ESMO 2024** retained TMB-H >= 10/Mb pan-tumor (tumour-agnostic, ESCAT IB). The tumour-type limits (poor performance in breast, prostate, glioma) come from **McGrail 2021**, not ESMO. ## Blood TMB (bTMB): The Negative-Trial Story **Gandara 2018** *Nat Med*: bTMB on Foundation Medicine FoundationACT panel; POPLAR + OAK retrospective. bTMB >= 16 mut/Mb showed PFS benefit with atezolizumab in NSCLC. **B-F1RST (Kim 2022)**: prospective phase 2 test of bTMB >= 16 as a first-line atezolizumab predictor in NSCLC; did NOT meet its pre-specified primary endpoint (bTMB-H improved ORR 28.6% vs 4.4%, only a non-significant PFS/OS trend). **BFAST Cohort C (Peters 2022)**: FAILED primary endpoint; atezolizumab vs chemo in bTMB-H NSCLC did not improve investigator-assessed PFS. Dominant confounder: low ctDNA shed fraction produces false-negative bTMB. **Operational state:** bTMB is research-grade in tissue-naive settings; tissue TMB remains the regulatory standard. ## Neoantigen Quality: Beyond Raw TMB **Luksza 2017** *Nature*: neoantigen fitness model. Combines "non-selfness" (TCR recognition probability via IEDB similarity) + "selfness" (MHC binding affinity differential vs WT peptide). Pancreatic-cancer validation (Balachandran 2017 *Nature*): long-term survivors had higher-quality neoantigens. Luksza 2022 *Nature*: immunoediting over 10 years. **McGranahan 2016** *Science*: **clonal neoantigen burden** (mutations present in all tumor cells) predicts ICI response better than total. Subclonal-rich tumors evade despite high TMB. **HLA-LOH** (McGranahan 2017 *Cell*, LOHHLA; Pyke 2022 *Nat Commun*, DASH; Montesion 2021 *Cancer Discov* for the ~17% pan-cancer estimate): HLA-LOH occurs in ~40% of NSCLC and abolishes neoantigen presentation for the lost allele. ~17% pan-cancer; >30% in HNSCC / NSCLC / cervical. Co-occurs with high subclonal burden + APOBEC + immune escape. ## Decision Tree by Scenario | Scenario | Recommended path | Why | |----------|------------------|-----| | Pan-tumor ICI eligibility (FDA pembrolizumab) | TMB-H >= 10/Mb on FoundationOne CDx | FDA companion diagnostic | | Non-FoundationOne panel | Apply per-assay calibration to the FoundationOne 10/Mb equivalent | TSO500 = 7.8; Oncomine = 8.4 (Ramos-Paradas 2021) | | WES TMB | Compute directly; threshold per ESMO 2024 = 10/Mb | WES is reference standard | | Tissue-naive bTMB | Caution: BFAST Cohort C failed | Research-grade; check ctDNA shed fraction | | Breast / prostate / glioma | TMB-H does not enrich ICI response per McGrail 2021 | Tumor-type-specific cutoffs | | MSI-H + TMB-H concurrence | MSI-H supersedes for ICI biomarker decision | Sha 2020 | | Hypermutator characterization (>=100/Mb) | Confirm MMR-D or POLE-exo via signatures + IHC | Co-occurrence is common | | Neoantigen quality (research) | Luksza fitness + HLA-LOH (LOHHLA / DASH) + clonality (McGranahan) | Beyond raw TMB | | Cross-panel comparison | Vega 2021 calibration equations OR WES re-projection | Direct comparison invalid | ## Standard Workflow **Goal:** Compute TMB from a VEP-annotated somatic VCF with full filtering. **Approach:** Parse cyvcf2; apply VAF + depth + germline (gnomAD) filters; count nonsynonymous coding consequences; divide by scored Mb. ```python from cyvcf2 import VCF import re NONSYNONYMOUS_CONSEQUENCES = { 'missense_variant', 'stop_gained', 'stop_lost', 'start_lost', 'start_retained', 'frameshift_variant', 'inframe_insertion', 'inframe_deletion', 'splice_donor_variant', 'splice_acceptor_variant', 'protein_altering_variant', 'initiator_codon_variant' } # Vega 2021-calibrated scored regions (Mb) PANEL_SCORED_REGION = { 'FoundationOne_CDx': 0.8, # Scored region; NOT 1.1 panel total 'MSK_IMPACT_v3': 0.98, 'MSK_IMPACT_v4': 1.22, 'TSO500': 1.3, # Scored from 1.94 total 'Oncomine_TML': 1.2, 'Caris_MI': 1.2, 'Tempus_xT_v3': 0.6, # Borderline reliability 'WES': 30.0, 'WGS': 3000.0 } # TMB2 (Ramos-Paradas 2021) equivalent thresholds for FDA 10/Mb FoundationOne sensitivity ASSAY_TMB_H_CUTOFF = { 'FoundationOne_CDx': 10.0, 'TSO500': 7.8, 'Oncomine_TML': 8.4, 'MSK_IMPACT_v3': 10.0, # Approximate; full Vega 2021 calibration recommended 'MSK_IMPACT_v4': 10.0, 'WES': 10.0 } def parse_consequences_from_vep(csq_field, csq_header): '''Parse VEP CSQ INFO field; returns list of per-transcript consequence types.''' if not csq_field: return [] cons_idx = csq_header.index('Consequence') out = [] for transcript in csq_field.split(','): fields = transcript.split('|') if len(fields) > cons_idx: out.append(fields[cons_idx]) return out def is_nonsynonymous(consequences, include_synonymous=False): '''Check if variant has nonsynonymous coding consequence. FoundationOne CDx convention INCLUDES synonymous (set include_synonymous=True). MSK-IMPACT and most academic pipelines exclude. ''' target = set(NONSYNONYMOUS_CONSEQUENCES) if include_synonymous: target.add('synonymous_variant') for cons_str in consequences: for cons in cons_str.split('&'): if cons in target: return True return False def calculate_tmb(vcf_path, scored_region_mb, csq_header, min_vaf=0.05, min_depth=100, max_gnomad_af=0.005, include_synonymous=False, exclude_hotspots=True, hotspot_bed=None): '''Calculate TMB with filtering per Vega 2021 harmonization. Args: scored_region_mb: panel's SCORED region (NOT total panel) min_vaf: 0.05 (FoundationOne) to 0.10 (tumor-only no UMI) max_gnomad_af: 0.005 (0.5%) typical for tumor-only germline filter include_synonymous: True for FoundationOne CDx-compatible; False for MSK-IMPACT exclude_hotspots: COSMIC drivers excluded (not random mutations) ''' vcf = VCF(vcf_path) cons_idx = csq_header.index('Consequence') if 'Consequence' in csq_header else 1 nonsyn_count = 0 total_pass = 0 for v in vcf: if v.FILTER is not None: # FILTER == None means PASS in cyvcf2 continue depth = v.INFO.get('DP', 0) if depth < min_depth: continue vaf = _get_vaf(v) if vaf is None or vaf < min_vaf: continue gnomad_af = v.INFO.get('gnomAD_AF', 0) or v.INFO.get('AF_popmax', 0) or 0 if gnomad_af > max_gnomad_af: continue total_pass += 1 csq = v.INFO.get('CSQ', '') consequences = parse_consequences_from_vep(csq, csq_header) if is_nonsynonymous(consequences, include_synonymous=include_synonymous): nonsyn_count += 1 tmb = nonsyn_count / scored_region_mb return { 'tmb': round(tmb, 2), 'nonsynonymous_count': nonsyn_count, 'total_passing_filters': total_pass, 'scored_region_mb': scored_region_mb } def _get_vaf(variant): '''Extract VAF from genotype FORMAT fields (Mutect2 AD or AF).''' try: ad = variant.format('AD') if ad is not None and len(ad) > 0: ad0 = ad[0] total = sum(ad0)