An explainable boosting machine model for identifying artifacts caused by formalin-fixed paraffin embedding

This study introduces FIFA, a novel, computationally efficient, and explainable boosting machine model that leverages local sequence context to effectively filter formalin-fixed paraffin-embedding (FFPE)-induced variant artifacts, thereby enhancing the reliability of retrospective cancer genomics research using archived samples.

The Gist

Imagine you have a massive library of old, dusty medical records. These records are stored in a special way called FFPE (Formalin-Fixed Paraffin-Embedded). Think of this like preserving a fruit in a jar of thick, sugary syrup to keep it from rotting for decades. It's great for long-term storage and is how most hospitals keep patient samples today.

However, there's a catch: the "syrup" (formalin) damages the fruit (the DNA) over time. When scientists try to read the genetic code from these old samples, the damage looks like typos. A "C" might look like a "T" just because the preservation process broke it, not because the patient actually had a mutation.

These "typos" are called artifacts. If a doctor or researcher mistakes a typo for a real disease-causing mutation, they might prescribe the wrong treatment or draw the wrong conclusions.

The Problem: Finding the Real Typos

Scientists have tried to fix this with various tools. Some are like simple spell-checkers that just say, "If a word appears less than 10% of the time, it's probably a typo." Others are like advanced AI spell-checkers that try to learn the context of the sentence.

The authors of this paper tested all these existing tools and found a frustrating truth: The fancy AI tools weren't always better than the simple ones, and they were often too complicated, expensive, or hard to update. It was like using a supercomputer to fix a typo in a text message.

The Solution: Introducing "FIFA"

The researchers built a new tool called FIFA (Filtering FFPE Artifacts).

Here is how FIFA works, using a simple analogy:

1. The Detective with a Magnifying Glass (Local Context)
Old tools looked at a single "typo" in isolation. FIFA is like a detective who doesn't just look at the word; they look at the whole neighborhood around it.

• Analogy: If you see the word "recieve" in a sentence, a simple check might miss it. But if you see it surrounded by other misspelled words and the handwriting looks shaky, you know it's a mistake. FIFA looks at the "neighborhood" of the DNA (the surrounding 500 letters) to see if the damage pattern looks like the "syrup" damage or a real mutation.

2. The Explainable Teacher (EBM Model)
FIFA uses a special type of AI called an Explainable Boosting Machine (EBM).

• Analogy: Most modern AI (like Deep Learning) is a "Black Box." You put data in, and it gives an answer, but it won't tell you why. It's like a teacher who says, "You got an A," but refuses to show you the grading rubric.
• FIFA is different. It's like a teacher who says, "You got an A because you used the right formula, your handwriting was clear, and your logic was sound." Because FIFA explains its reasoning, scientists can trust it and tweak it if they need to.

3. The Lego Builder (Easy Updates)
One of the biggest problems with old AI tools is that if you get new data, you have to rebuild the whole thing from scratch.

• Analogy: FIFA is built like Lego blocks. If a new hospital sends in a new batch of samples, you don't have to tear down the whole castle. You just snap a new Lego block onto the existing structure. This makes FIFA incredibly easy to update as new data becomes available.

Why This Matters

The researchers tested FIFA on thousands of samples from different types of cancer (lymphoma, breast cancer, etc.). They found that:

• It works better: FIFA caught more real mutations and filtered out more fake ones than the other tools.
• It's fast and cheap: You don't need a supercomputer to run it; a standard laptop can handle it.
• It reveals the truth: When they used FIFA to clean up the data, the biological signals (like specific patterns of cancer mutations) became much clearer. It was like cleaning a dirty window; suddenly, you could see the view outside perfectly.

The Bottom Line

FFPE samples are a goldmine of medical history, but they are covered in "dust" (artifacts) that makes them hard to read. FIFA is a new, smart, and transparent tool that sweeps that dust away. It helps doctors and researchers see the real genetic story behind the samples, potentially leading to better treatments and cures, all without needing expensive, complicated equipment.

It turns a dusty, confusing archive into a clear, readable story.

Technical Summary

1. Problem Statement

Formalin-fixed paraffin-embedding (FFPE) is the standard method for long-term clinical sample storage, with over 400 million archived samples globally. However, formalin fixation chemically modifies nucleic acids, causing fragmentation, crosslinks, and deamination (manifesting as artifactual C>T mutations). These artifacts complicate somatic variant calling (sSNV), often leading to false positives that obscure true biological signals.

Existing computational methods to filter these artifacts face several limitations:

• Simple Thresholds: Variant Allele Frequency (VAF) cutoffs (e.g., <10%) are easy to implement but fail in complex scenarios involving low tumor purity or subclonality, where real variants also have low VAFs.
• Machine Learning (ML) Limitations: Tools like FFPolish (logistic regression) and Ideafix (XGBoost) incorporate read-level features but often ignore the broader local genomic context surrounding a variant.
• Deep Learning (DL) Limitations: Tools like DeepSomatic (CNN-based) perform well but are "black boxes," difficult to interpret, computationally expensive (requiring GPUs), and hard to retrain or update with new data.
• Generalizability: Many existing models struggle to generalize across different cohorts due to batch effects and varying fixation protocols.

2. Methodology

The authors developed FIFA (Filtering FFPE Artifacts), a novel tool based on an Explainable Boosting Machine (EBM).

Data Collection and Labeling

• Training Data: Utilized 90 paired Fresh Frozen (FF) and FFPE samples from four cohorts: NYGC1 (13 DLBCL), NYGC2 (7 B-cell lymphoma), BLGSP (Burkitt Lymphoma), and HTMCP (Cervical Cancer).
• Ground Truth: Variants called in FFPE were labeled "Real" if supported by any evidence in the matched FF sample; otherwise, they were labeled "Artifacts."
• Independent Validation: Tested on HCC1395 breast cancer cell line data (SEQC2) and an independent breast cancer cohort (NYGC3, 143 samples).

Feature Engineering

FIFA utilizes a set of 60 features per variant, combining established metrics with novel context-aware features:

1. Standard Features: VAF, mapping quality, base quality, proximity to read ends, and fragment length (shared with FFPolish and Ideafix).
2. Local Context Features: To emulate the performance of image-based CNNs without the computational cost, FIFA extracts features from a ±500bp window flanking the variant. This includes median fragment length, duplicate fraction, and improper read pairing rates within the window.
3. MOBSTER Probability: Repurposed the MOBSTER tool to calculate the probability of a variant lying in the "neutral tail" (low-frequency passenger mutations), using this as a proxy for FFPE artifact likelihood.
4. Transformation: Numerical features were transformed using an inverse hyperbolic sine function to handle zero/negative values and normalize distributions across batches.

Model Architecture: Explainable Boosting Machine (EBM)

• Algorithm: EBM is an ensemble of generalized additive models (GAMs) using decision trees.
• Advantages:
  • Interpretability: Provides global and local feature importance without needing post-hoc attribution methods (like SHAP).
  • Efficiency: Fast training and inference on standard CPUs.
  • Modularity: Models trained on different cohorts can be "averaged" (merged) to create a consolidated model, allowing for easy updates as new data becomes available.
• Training: Hyperparameters were optimized using Optuna (5-fold cross-validation) on each cohort independently.

3. Key Results

Benchmarking Existing Methods

• In initial evaluations on lymphoma cohorts, sophisticated ML methods (Ideafix, FFPolish) and Deep Learning (DeepSomatic) showed highly variable performance.
• Surprisingly, a simple 10% VAF cutoff often outperformed complex ML methods in terms of F1 score and accuracy, particularly in the NYGC2 cohort, highlighting the difficulty of generalizing complex models without robust training data.
• DeepSomatic performed well on NYGC1 but suffered a significant drop in sensitivity on NYGC2.

FIFA Performance

• Round-Robin Validation: When trained on three cohorts and tested on the fourth (leave-one-cohort-out), FIFA consistently achieved the highest F1 scores (ranging from 0.81 to 0.95) across NYGC1, NYGC2, and HTMCP, outperforming FFPolish and VAF cutoffs.
• HCC1395 Cell Line: On the independent SEQC2 dataset, FIFA achieved a mean F1 score of 0.961, significantly outperforming DeepSomatic (0.88–0.93) and FFPolish. Notably, FIFA applied as a post-hoc filter to standard Mutect2 calls achieved performance comparable to the end-to-end DeepSomatic pipeline.
• Biological Validation (NYGC3):
  • RNA-seq Corroboration: In 143 breast cancer samples with matched RNA-seq, FIFA showed higher precision in retaining variants supported by RNA reads compared to other methods.
  • Mutational Signatures: Filtering with FIFA increased the similarity of mutational signatures to known breast cancer profiles (ICGC/GEL) while reducing similarity to unrelated cancers (colorectal/bone).
  • HRD Detection: FIFA filtering significantly increased the proportion of the SBS3 signature (associated with Homologous Recombination Deficiency/BRCA mutations) in HRD-positive samples, indicating better preservation of biologically relevant signals.

4. Key Contributions

1. Novel Tool (FIFA): A pipeline-agnostic, post-hoc filtering tool that uses EBM to balance high performance with interpretability.
2. Context-Aware Features: Demonstrated that capturing local genomic context (via window-based features) is critical for distinguishing artifacts, a feature often missing in traditional ML tools but present in DL models.
3. Modular Model Updating: Introduced a workflow where independent EBM models can be merged ("averaged") to incorporate new cohorts without retraining from scratch, solving the data scarcity and batch effect issues common in FFPE research.
4. Accessibility: Unlike DL models, FIFA requires modest computational resources (runs on standard CPUs, <2.75GB RAM) and is freely available for academic use.

5. Significance

This study addresses a critical bottleneck in cancer genomics: the reliable utilization of the vast archive of FFPE clinical samples. By proving that an interpretable, lightweight model can outperform or match state-of-the-art deep learning methods, the authors democratize access to high-fidelity variant calling.

FIFA enables researchers to:

• Retrospectively analyze FFPE samples with higher confidence.
• Detect low-frequency subclonal mutations that simple VAF cutoffs would miss.
• Extract accurate mutational signatures and biological signals (e.g., HRD status) essential for precision oncology.

The authors conclude that FIFA represents a significant advance in making the "vast stores of FFPE-archived tumor samples" a viable resource for future clinical and research discoveries.