TL;DR
This paper introduces Distribution-Aware In-Context Learning (DA-ICL), a framework that moves beyond treating genetic ancestry as a set of discrete categories (e.g., "African" vs. "European"). By leveraging tabular foundation models and a novel context-enhancement technique, the authors demonstrate that modeling ancestry as a continuous spectrum significantly improves cancer outcome predictions, especially for individuals who fall between traditional ancestral "centroids."
What Problem Does This Paper Address?
Most machine learning models in genomics follow an ancestry-discrete design: they either ignore ancestry entirely or partition patients into broad, categorical groups. However, human genetics exists on a continuum. Recent studies have shown that predictive accuracy for polygenic scores and clinical outcomes decays linearly as a patient's genetic distance increases from the training set's center—even within a single "ancestry group."
The challenge is that the relationship between omics features (like gene expression) and clinical endpoints (like survival) is nonstationary; it shifts continuously across the genetic landscape. Traditional models fail to capture these subtle shifts, leading to performance disparities for admixed populations or those underrepresented in biobanks.
Methods
The authors propose a two-pronged approach to ancestry-continuum modeling:
- Efficient Modeling via In-Context Learning (ICL): Instead of training a new model for every point on the genetic spectrum, the authors use TabPFN, a pretrained tabular foundation model. At inference time, they provide the model with a "context"—a small set of reference samples that are genetically similar to the target patient. The model adapts its predictions to this local context without any parameter updates.
- Distribution-Aware Context Enhancement (DACE): A local neighborhood in genetic space might be data-sparse. DACE enriches this context by searching the entire dataset for "informative samples." It uses a nonparametric $f$-divergence estimation method based on Minimum Spanning Trees (MST). By identifying samples that lie on the "cut edges" between the local neighborhood and the rest of the data, DACE finds patients who are distributionally similar in their omics-endpoint relationships, even if they are genetically more distant.
Datasets
The framework was validated using both synthetic and real-world data:
- Synthetic Data: A 2D ancestry space with three noisy clusters connected by admixed samples, where the decision boundary for the outcome variable was a function of the ancestry coordinates.
- TCGA (The Cancer Genome Atlas): Data from three cancer types—Breast (BRCA), Kidney (KIRC), and Uterine (UCEC).
- Omics Modalities: mRNA expression, microRNA expression, and DNA methylation.
- Endpoints: Overall Survival (OS) and Progression-Free Interval (PFI).
Key Insights and Ideas
- Theoretical Superiority: The authors provide a mathematical proof showing that a Bayes-optimal ancestry-continuum classifier will always have an error rate less than or equal to a Bayes-optimal ancestry-discrete classifier. Essentially, discarding within-group heterogeneity is always suboptimal.
- Zero-Shot Adaptation: By using ICL, the system can handle the "ancestry-continuum" problem without the massive computational overhead of retraining or fine-tuning models for every possible ancestral mix.
- Distributional Proximity > Genetic Proximity: DACE recognizes that two patients might have different genetic backgrounds but similar omics-to-outcome mappings. Finding these "distributional neighbors" is key to overcoming data sparsity in specific regions of the genetic continuum.
Key Results
- Synthetic Validation: In simulations, as the relationship between features and outcomes became more dependent on ancestry (higher rotation/offset), the continuum models significantly outperformed discrete models.
- Real-World Performance: Evaluated on African (AFR) test samples in TCGA, DA-ICL achieved the highest median AUROC in 11 out of 18 tasks compared to standard baselines like XGBoost, Random Forest, and Elastic Net.
- Robustness to Genetic Distance: While traditional models like XGBoost showed rapid performance decay as patients moved further from the ancestral group centroid, DA-ICL maintained relatively stable performance across the continuum.
Limitations and Future Work
- Finite Sample Constraints: While the theory proves continuum modeling is superior, the authors note that in practice, performance depends on having enough local or "informative" samples to build a high-quality context.
- Task Dependency: The benefits of DACE varied across tasks, suggesting that the optimal way to "enhance" context might depend on the specific cancer type or omics modality.
- Future Directions: The authors suggest exploring more complex genetic spaces beyond 30 PCA components and investigating how this framework scales to even larger, more diverse biobanks.
Summary updated