Finding cancer mutations in sequenced tumor DNA works best when several detection programs ("variant callers") are run side by side and a learned classifier, a meta-caller, decides which candidate mutations to believe. The strongest meta-callers are boosted decision trees trained on labeled examples. They learn which combinations of read depth, strand bias, mapping quality and caller agreement signal a real mutation rather than a sequencing artifact. These classifiers are only as good as their labels, and the labels come from a handful of public reference samples whose sequencing conditions look little like a real hospital's. Hospitals cannot pool their patients' sequencing reads to fix this. They also usually have no ground truth for their own samples.
We propose that each hospital manufacture its own labeled training data by computationally inserting known synthetic mutations ("spike-ins") into its own real sequencing reads. This captures that site's library chemistry, depth, and tumor purity. The hospitals would then jointly train one gradient-boosted meta-caller by federated learning, meaning they exchange only aggregated gradient statistics and never reads or per-variant features. The two-sentence version: local spike-ins give every site realistic labels for free, and federated boosting lets the shared model learn conditions that any single site sees too rarely, such as low-purity tumors or subclonal mutations at 5–10% allele frequency. The claim we will test is narrower than "federation helps." Iterative federated tree-building should beat both local-only training and simply averaging locally trained models, and the advantage should sit specifically in sequencing conditions that are sparse at every individual site.
Where this sits
Supervised ensemble meta-calling is established and centralized. SomaticSeq combines five SNV callers and three indel callers with 70+ per-site features in an AdaBoost classifier. On the high-contamination DREAM Setting D it reaches F1 90.5% against 62.4% for the best single tool, MuTect. SICaRiO trains XGBoost on 219 reference-derived features for indel filtering. Hofmann et al. show two things in their exome simulations: a rank-combination of five callers reaches 78% sensitivity at 90% precision, against 71% for the best single tool, and simple intersection misses nearly 60% of true mutations. These papers supply our feature set and model family. They also motivate the proposal, because all of them depend on co-located truth sets, and the file lists that dependence as the main barrier to generalizing to clinical cohorts.
Protocol heterogeneity is the reason local labels matter. Alioto et al. found large discordance across 18 somatic pipelines. They also found that PCR-free libraries reach 73–74% coverage at ≥25x while PCR-based libraries drop as low as 46%. Yi et al. show, centrally, that giving a neural caller an explicit sequencing-type input (SeqType) reduces indel error by 6% to 38% when mixing WGS and WES. We borrow that idea in tree form by making protocol descriptors ordinary input features. HELLO uses a similar context-gated combination of platform experts. It is centralized and requires co-located read libraries.
On the federated side, SimFL trains GBDTs across horizontal silos using locality-sensitive hashing (a way to find similar records across sites without sending them) plus similarity-weighted gradients. It outperforms local-only training by an average of 4% on LIBSVM data but has never touched genomic or heavily imbalanced data. Sadilek et al. show that federated clinical models can match centralized ones: heart-failure AUC was 0.85 federated versus 0.82 in the original study. FedBEns and FedBE represent the one-shot and ensemble-style alternatives that we treat as serious competitors.
Genomic federated learning to date covers association studies (dynamic-sampling GWAS, biobank-scale secure GWAS) and expression classifiers (federated tumour classification). All of these take already-called variants as fixed inputs. None trains the step that decides what a variant is.
SMC-Het constrains our scope. In that challenge, ensembles failed to beat the best individual subclonal reconstruction method. We therefore restrict the claim to mutation calling, where learned ensembles have a documented margin, and we do not assume ensembling helps downstream reconstruction.
Core Claim
Each site can generate labels by running a spike-in tool such as BAMSurgeon (the simulator used in SMC-Het) on its own real normal or tumor reads, at controlled allele frequencies and purities. A meta-caller trained on these labels captures site-specific artifacts that public truth sets miss. Federated histogram-based gradient boosting over these local labels, with protocol descriptors as features, should approach a hypothetical pooled model. We also predict that the one-shot combination of local models will fall measurably short specifically in condition strata that are rare at every site. The reason is mechanical. A tree split in a sparse stratum needs enough gradient mass to be chosen. Federated boosting sums that mass across sites before choosing the split, whereas each local model sees too little of it to ever make that split.
Research Program
Each site runs the SomaticSeq caller panel on its own data, including spike-in versions of its own samples, and computes SomaticSeq-style features. We deliberately exclude genomic coordinates and alleles from the feature set, which limits re-identification risk.
Training uses a federated XGBoost-style protocol. In every round, each site computes gradient and Hessian histograms over shared quantile bins. The server sums these histograms under secure aggregation, so it sees only the totals, and chooses the tree split. Optionally, calibrated noise is added to the histograms for differential privacy. We compare this protocol against SimFL's hashing approach. We favor histograms a priori because per-instance hashes of variant features may leak more than aggregate statistics, and we will measure the communication cost of each.
Three analyses follow the training. First, we check label realism by comparing the error profile of spike-in-trained models against models trained on expert-curated real truth. Second, we test whether the federated advantage concentrates in sparse strata, stratifying results by allele frequency, purity, repeat context and protocol. Third, we test whether protocol descriptors let a single shared model serve heterogeneous sites, or whether per-site calibration of the output threshold is still needed. SomaticSeq's optimal cutoff of P ≥ 0.7 is a single global number, and we expect it to move across sites.
Baselines
The naive reference points are the best single caller, MuTect, and majority vote over the same caller panel, run with default settings on each site's data. SomaticSeq's DREAM result and Hofmann's intersection result predict that both will be weak. They anchor the scale of improvement.
The strongest existing method is SomaticSeq trained centrally on public truth and deployed unchanged at every site. It uses the reported 90.5% F1 (DREAM Setting D) as a sanity check that our reimplementation reproduces it. On that same benchmark it must be matched before any cross-site claim counts.
Local-only training (SOLO in SimFL's terms) fits the same booster to each site's own spike-ins with an identical tuning budget: the same number of trees, the same depth grid, and the same early-stopping rule. SimFL reports that its federated method beats SOLO by an average of 4% on LIBSVM. For variant calling, no reported result was found on Lacuna for federated meta-callers.
One-shot aggregation has each site train locally and send its finished model. The server averages predicted probabilities or applies a FedBEns-style Bayesian combination. This baseline isolates whether iterative communication matters at all.
Pooled-centralized training on all sites' spike-ins is the upper bound, not a competitor. Two ablations complete the set: removing the protocol-descriptor features, and replacing local spike-in labels with public-truth labels while keeping federation.
Beating the field means two things. On held-out real truth, federated training must exceed both public-trained SomaticSeq and local-only training in sensitivity at 90% precision. It must also exceed one-shot aggregation within the rare strata, with confidence intervals from bootstrapping over samples.
Why not just…?
Why not just deploy SomaticSeq trained on public benchmarks? Public training sets are high-purity cell lines or synthetic tumors, and Alioto et al. show that protocol alone reshapes coverage and error. The public-trained baseline is in our comparison. If it matches the local-label models on the curated ICGC truth, the spike-in half of the proposal fails, and we will report that.
Why not just train on each site's own spike-ins with no federation? If a site has abundant data in every condition, this may be sufficient. Our prediction is that it fails only in strata that are sparse at every site. The stratified evaluation tests this directly, and a null result would mean federation is unnecessary for well-resourced sites.
Why not just share finished local models and average them? This is the most serious threat. It is cheap and needs no communication rounds, and FedBEns-style methods exist to make it work. We made it a primary baseline and designed the site partition so that some conditions are thin everywhere, which is exactly where our mechanism predicts one-shot aggregation fails. If it matches iterative boosting even there, we will recommend it instead.
Why not just share de-identified feature tables centrally? Per-candidate features at known loci remain linkable to individuals, a point raised by Ching et al.. We cannot test legal feasibility. The pooled model does, however, tell us exactly how much accuracy federation gives up relative to sharing.
Proposed Experiments
We build a simulated consortium from public data so that real truth exists for evaluation. The virtual sites are drawn from the ICGC CLL and medulloblastoma benchmark. Its pipelines from different centers and its PCR-free and PCR-based libraries become different sites, and we downsample from 20x to 250x. Its expert-curated truth at roughly 300x depth serves as the real-truth test set. Additional sites come from the DREAM synthetic settings and from Hofmann's exome framework, which spans 12x–212x depth and 10–60% contamination. The HG008 tumor-normal pair serves as a held-out Illumina site. It has no reported caller benchmarks, so we will report its numbers as new.
Each site trains only on spike-ins into its own reads. All evaluation uses real curated truth, with leave-one-site-out testing. Metrics are F1 and sensitivity at 90% precision, both stratified by allele frequency (with the 5–10% bin singled out), purity and repeat context. We also report bytes communicated and number of rounds.
Four checks decide the proposal. First, spike-in-trained models must match or beat public-truth-trained models on real ICGC truth. Second, federated training must close most of the gap between local-only and pooled training. Third, the advantage of federated over one-shot aggregation must grow as stratum sparsity increases. Fourth, removing the protocol descriptors must hurt most on the PCR-based and exome sites.
References
- An ensemble approach to accurately detect somatic mutations using SomaticSeq
- SICaRiO: Short Indel Call filteRing with bOosting
- Detailed simulation of cancer exome sequencing data reveals differences and common limitations of variant callers
- A comprehensive assessment of somatic mutation detection in cancer using whole-genome sequencing
- Learning from Data-Rich Problems: A Case Study on Genetic Variant Calling
- HELLO: A hybrid variant calling approach
- Practical Federated Gradient Boosting Decision Trees
- Privacy-first health research with federated learning
- FedBEns: One-Shot Federated Learning based on Bayesian Ensemble
- FedBE: Making Bayesian Model Ensemble Applicable to Federated Learning
- Privacy-preserving federated genome-wide association studies via dynamic sampling
- Secure and federated genome-wide association studies for biobank-scale datasets
- Differentially Private, Federated Learning on Gene Expression Data for Tumour Classification
- Crowd-sourced benchmarking of single-sample tumor subclonal reconstruction
- Opportunities and obstacles for deep learning in biology and medicine
- Development and extensive sequencing of a broadly-consented Genome in a Bottle matched tumor-normal pair