Greytak et al. (2015) — Accuracy of Molecular Data Generated with FFPE Biospecimens
Bibliographic Reference
Greytak, S. R., Engel, K. B., Bass, B. P., & Moore, H. M. (2015). Accuracy of molecular data generated with FFPE biospecimens: Lessons from the literature. Cancer Research, 75(8), 1541–1547. DOI: 10.1158/0008-5472.CAN-14-2378.
Core Argument
FFPE biospecimens are the largest archival resource for cancer molecular research, but the genomic and expression data derived from them may not accurately reflect patient physiology — they reflect a composite of true biology, formalin-induced artifacts, and processing variables. Through a systematic survey of 34 representative articles comparing case-matched FFPE and fresh/frozen human cancer biospecimens, the authors demonstrate that concordance varies dramatically by platform, gene, transcript, preservation protocol, and analytical parameter. The central message: validation of FFPE-optimized protocols against case-matched frozen controls is mandatory for each platform, gene, and fixation regime — concordance rates alone mask clinically consequential differences in false discovery rates, copy number, and transcript-specific effects.
Methods
- Data source: NCI Biospecimen Research Database (BRD) + PubMed cross-referencing
- Initial screen: 68 primary research articles comparing genomic/expression endpoints in case-matched FFPE vs. fresh/frozen human neoplastic tissue
- Final set: 34 representative articles selected by: published within last 10 years, ≥5 patients, current/relevant platforms. Exceptions for novel findings.
- Endpoints surveyed: Genotyping (SNV, indel), copy number, DNA methylation, RNA expression (mRNA, miRNA) by qRT-PCR, microarray, DASL, and NGS
- Analysis: Qualitative synthesis organized by analyte (DNA, RNA) and platform; no formal meta-analysis due to heterogeneity of methods and endpoints
Key Findings
-
Genotype concordance is platform-dependent but universally imperfect. NGS achieved >99% concordance with frozen controls at 40× coverage, while traditional Sanger sequencing ranged from 59–82%. However, even the best platforms generated false-positive SNVs: NGS false discovery rates ranged from <1% to 15% depending on coverage depth (5–80×). KRAS — a clinically actionable gene — showed 6–20% misclassification rates in FFPE vs. frozen comparisons across multiple studies. Increasing NGS coverage from 20× to 40× reduced discordant loci from 1% to 0.2%.
-
FFPE introduces characteristic mutational artifacts. Compared to frozen controls, FFPE biospecimens showed elevated rates of transversion mutations, transition mutations, and small indels. Mutations at A:T base pairs were more common in FFPE than G:C mutations, but GC-content of the target sequence was the dominant factor: the strongest FFPE-frozen correlation occurred at 40% GC-content (r=0.97), and SNP detection was most successful in the 35–55% GC range. Artifactual C>T mutations from cytosine deamination — the same chemistry underlying APOBEC mutagenesis but occurring ex vivo during fixation — are the hallmark FFPE artifact.
-
Copy number analysis is unreliable without validation. Published studies conflicted on whether FFPE-frozen CNA concordance is good, poor, or platform-dependent. Whole-genome amplification reduced concordance from 59% to 48%. The largest disparities were platform-driven: the same biospecimens showed >98% agreement on Agilent 4×44K arrays but only 53.8–87.3% on Affymetrix SNP 6.0 arrays. Tumor heterogeneity (stromal admixture) was identified as an additional confounder beyond fixation effects.
-
RNA expression: the diagnosis × preservation interaction is alarming. mRNA expression correlations between FFPE and frozen ranged from r=0.02 to r=0.96 depending on platform and protocol. Critically, a statistical interaction between diagnosis and fixation method was identified: only 33% of genes differentially expressed between tumor and normal in FFPE were also differentially expressed in frozen, and only 48% of frozen-identified DEGs were confirmed in FFPE. Affected transcripts included breast cancer genes APC, CDKN2A, IGF1R, TGFA, TSG101, and ESR1. miRNA was substantially more stable (r >0.94 in some studies), clustering by diagnosis rather than preservation method.
-
Key modifiable factors for FFPE accuracy. GC-content of the target sequence (optimal 40–60%), probe location within transcript (3’ UTR outperforms 5’), amplicon size (≤105 bp for qRT-PCR), number and choice of normalization transcripts, use of transcript repair protocols, and platform selection (exon arrays outperform U133 arrays; NGS outperforms traditional sequencing) all substantially affect FFPE-frozen concordance.
Concepts Introduced or Used
FFPE, formalin fixation, biospecimen pre-analytical variables, cold ischemia time, fixation duration, GC-content bias, C>T deamination artifact, diagnosis × preservation interaction, false discovery rate, copy number concordance, transcript repair, NGS coverage depth, microarray platform effects, miRNA stability
Entities Referenced
- KRAS — clinically actionable gene showing 6–20% FFPE-frozen misclassification
- APC, CDKN2A, IGF1R, TGFA, TSG101, ESR1 — breast cancer genes with diagnosis × preservation interaction effects
- NCI Biospecimen Research Database (BRD) — source of the systematic review
- Affymetrix SNP 6.0, Agilent 4×44K, U133 Plus 2.0, Human Exon 1.0 ST — microarray platforms with differing FFPE performance
- NuGAN FFPE labeling system
Limitations (as stated by authors)
- Reliance on concordance as the primary metric is itself a limitation — correlations fail to capture clinically relevant differences in false discovery rates, copy number disparities, and transcript- or promoter-specific effects
- The literature survey was limited to 34 representative articles; a formal meta-analysis was precluded by heterogeneous methods and endpoints
- Most studies used biospecimens from a single institution with uniform fixation protocols; results may not generalize across different FFPE processing regimes (especially fixation duration >72h)
- Whether optimization of fixation and processing protocols (rather than post-hoc analytical optimization) would improve data reliability remains undetermined
- Formalin-free fixatives (BHP, PAXgene, etc.) show promise but are not yet widely adopted
Relevance to Clonal Evolution
This paper is the foundational reference for understanding why pre-analytical variables constrain every inference made from FFPE-derived cancer genomics data. For the clonal evolution framework, the implications cascade through the entire measurement pipeline:
- VAF → CCF translation assumes the VAF reflects tumor biology, not fixation artifact. FFPE-induced C>T mutations at low allele fractions can be misclassified as subclonal mutations, inflating SMF and creating phantom subclones. The 1–15% false discovery rate for NGS SNVs in FFPE directly translates to a 1–15% phantom subclonal mutation fraction.
- CNA inference from logR/BAF assumes signal reflects copy number, not GC-content bias. The platform-dependent CNA discordance (53–98%) means that copy-number calls in FFPE-derived data carry an additional layer of uncertainty beyond the biological and computational confounders already documented in copy-number-alteration.
- The diagnosis × preservation interaction means that differentially expressed genes identified from FFPE tumor/normal comparisons may be artifacts of fixation susceptibility rather than cancer biology. For studies linking transcriptional ITH to clonal architecture, this interaction is a confounder that cannot be removed post-hoc — it must be prevented by validation against frozen controls.
The paper establishes that FFPE is not a neutral substrate — it introduces systematic biases that are gene-specific, transcript-specific, platform-specific, and diagnosis-specific. These biases propagate into every downstream evolutionary inference: clone detection, phylogenetic reconstruction, mutational signature extraction, and differential expression analysis.