What limits local ancestry inference at low divergence: a feasibility threshold, a metric that conceals failure, and a deficit of input more than architecture

Technical Summary: What Limits Local Ancestry Inference at Low Divergence

Problem Statement
Local ancestry inference (LAI) assigns genomic positions in admixed individuals to source populations, a prerequisite for admixture mapping and ancestry-specific association testing. While routine LAI methods perform well on highly divergent sources (e.g., African/European/Native American, $F_{ST} \approx 0.1$), their efficacy on closely related populations (e.g., Northern vs. Southern Han Chinese, $F_{ST} \approx 0.0004$) remains untested. Current benchmarks rely heavily on coalescent simulations and per-site accuracy metrics, which may not reflect the tract-level structure required for downstream analyses like admixture dating. This study investigates the feasibility limits of LAI at low divergence, the validity of simulation-based benchmarks, and the relative contributions of input representation versus model architecture.

Methodology
The author evaluated five methods across a continuous gradient of source divergence ($F_{ST}$ from 0.0022 to 0.243) using two distinct data tracks:

1. Coalescent Simulations: Two-way admixture events were simulated with exact ground truth labels generated by mosaicking donor haplotypes.
2. Real Haplotypes: The same mosaicking construction was applied to phased 1000 Genomes haplotypes across 11 population pairs (spanning East Asian, European, and South Asian groups).

Methods Compared:

• Baselines: A windowed likelihood classifier and a likelihood classifier smoothed by a Hidden Markov Model (HMM).
• Released Tools: RFMix v2 and FLARE.
• Neural Network: A dilated convolutional neural network (CNN) trained for per-site segmentation. The network was tested in two configurations:
  • Frequency-only: Input consists of allele frequencies and log-likelihood ratios.
  • Haplotype-aware: The same architecture augmented with four channels summarizing local haplotype matching against reference panels.

Key Contributions and Findings

1. A Feasibility Floor Exists
There is a hard limit to LAI performance determined by the information content of the data rather than algorithmic sophistication.

• At $F_{ST} = 0.0022$, no method exceeds 0.575 accuracy.
• For the CHB/CHS pair ($F_{ST} = 0.00042$), the best accuracy is 0.551.
• Below these thresholds, differences between methods are smaller than run-to-run variability, indicating that the data contains insufficient signal to distinguish sources. This implies that published inferences of fine-scale ancestry tracts within such populations (e.g., within Han Chinese or within Europe) lack support under current conditions.

2. Per-Site Accuracy is a Misleading Metric
The standard metric, per-site accuracy, fails to capture the structural integrity of the inferred mosaic.

• The dilated CNN achieved high per-site accuracy in simulations but produced 78.8 times more ancestry tracts than the truth, implying an admixture time 61.2 times older than reality.
• In contrast, RFMix and FLARE produced tract counts close to the truth.
• Applying a Viterbi decoder (using a fixed transition prior based on recombination) to the CNN's logits corrected the tract structure (reducing the error to 1.56×) with negligible cost to per-site accuracy (+0.0002). This demonstrates that the metric is blind to defects that render outputs unusable for admixture dating.

3. Simulation Does Not Predict Real-World Performance
A learned method's advantage in simulation does not transfer to real data.

• In simulations, the dilated CNN outperformed the best released tool by up to 0.048 at low divergence.
• On real 1000 Genomes haplotypes, the CNN was outperformed by released tools on 10 of 11 pairs, with a mean deficit of -0.032.
• This discrepancy is not due to the simulator generating "cleaner" data; simulated panels actually carried more exploitable haplotype signal than real panels at matched divergence. The failure arises because the frequency-only network relies on information that is insufficient on real chromosomes, whereas released tools utilize full haplotype information.

4. Input Representation is the Primary Bottleneck
The performance gap is driven by input representation rather than architectural choices.

• Architectural Interventions: Variations in attention mechanisms, state-space layers, capacity, and pretraining moved accuracy by at most 0.006.
• Input Interventions: Supplying the CNN with haplotype-matching channels (matching the input of RFMix/FLARE) recovered +0.031 accuracy on 8 of 8 pairs below $F_{ST} = 0.04$.
• However, even with this input parity, the network remained behind the released tools on 10 of 11 pairs, suggesting that while input is necessary, it is not sufficient to close the gap entirely. The network only exceeded released tools on the CEU/TSI pair ($F_{ST} = 0.0027$), the least divergent pair where any method was informative.

5. Reference Panel Size and Statistics Matter More Than Architecture
Two factors conventionally held fixed in comparisons were found to be more influential than architectural design:

• Summary Statistic: Neural methods typically use agreement rates (Hamming distance), while copying models (RFMix/FLARE) use contiguous agreement lengths (sufficient statistics for the Li–Stephens model). Contiguity is more discriminative but only provides a benefit when the reference panel is small; it acts as a substitute for panel size rather than an independent signal.
• Panel Size: No method approached saturation. Increasing the reference panel from 80 to 100 haplotypes improved accuracy for all methods (including the CNN, RFMix, and FLARE) without changing their relative ordering.

6. Sensitivity to Phasing Errors
Phase switch errors fragment haplotypes without altering allele frequencies.

• Methods relying on long-range integration (like the frequency-only CNN) are most sensitive to these errors, suffering greater accuracy loss than methods that do not integrate context (windowed likelihood) or those that explicitly model haplotypes (RFMix/FLARE).
• Adding haplotype-matching channels reduced this sensitivity, making the haplotype-aware configuration more robust to phasing errors than the frequency-only version.

Significance and Claims
The paper argues that the limitations of LAI at low divergence are primarily data-driven (insufficient signal in the genome) and representational (mismatched input information), rather than a failure of learned inference architectures. It highlights three specific pitfalls in current evaluation practices:

1. Relying on per-site accuracy, which masks structural failures critical for downstream analysis.
2. Validating solely on simulations, which can artificially inflate the performance of learned methods compared to released tools.
3. Comparing methods with unequal input representations (e.g., frequency-only vs. haplotype-aware).

The author concludes that for closely related populations, the "floor" of performance is determined by the information content of the data. While supplying haplotype information improves performance, it does not guarantee parity with established tools, and claims of fine-scale ancestry inference in these regimes require independent support beyond the inference itself. The study suggests that future benchmarks must report tract-level statistics, validate on real haplotypes, and ensure input parity between compared methods.