SwiftRepertoire: Few-Shot Immune-Signature Synthesis via Dynamic Kernel Codes

TL;DR

SwiftRepertoire tackles label sparsity in TCR repertoire diagnostics by learning geometry-preserving prototypes and synthesizing sparse, task-specific adapters conditioned on compact descriptors. The adapters are produced through a constrained proximal retrieval process and applied to a frozen backbone, enabling rapid, few-shot adaptation with interpretability via motif-aware probes and calibrated motif testing. The approach is supported by formal statistical tests (Fisher energy ratio, bootstrap coverage) and a two-stage motif calibration pipeline, with Neural-ODE components for continuous-time modeling where beneficial. Empirical results on cancer datasets show strong few-shot performance, robust generalization, and practical potential for low-resource clinical deployment, underpinned by theoretical guarantees linking geometry, prototype coverage, and generalization risk.

Abstract

Repertoire-level analysis of T cell receptors offers a biologically grounded signal for disease detection and immune monitoring, yet practical deployment is impeded by label sparsity, cohort heterogeneity, and the computational burden of adapting large encoders to new tasks. We introduce a framework that synthesizes compact task-specific parameterizations from a learned dictionary of prototypes conditioned on lightweight task descriptors derived from repertoire probes and pooled embedding statistics. This synthesis produces small adapter modules applied to a frozen pretrained backbone, enabling immediate adaptation to novel tasks with only a handful of support examples and without full model fine-tuning. The architecture preserves interpretability through motif-aware probes and a calibrated motif discovery pipeline that links predictive decisions to sequence-level signals. Together, these components yield a practical, sample-efficient, and interpretable pathway for translating repertoire-informed models into diverse clinical and research settings where labeled data are scarce and computational resources are constrained.

Figures (8)

• Figure 1: Overview of the SwiftRepertoire pipeline. The figure summarizes the main stages and diagnostics used in SwiftRepertoire: low-dimensional adapter diagnostics (PCA energy and Fisher summaries), formal Fisher-energy hypothesis testing (bootstrap + Bonferroni correction), prototype construction and canonicalization, and the constrained proximal retrieval training with nested anti-leakage partitions and diagnostics. Neural-ODE components with stiff-stable integrators are integrated into the retrieval and descriptor pipelines to provide continuous-time modeling where appropriate.
• Figure 2: t-SNE projection of adapter vectors and learned prototypes. Points are colored by disease / task ID; black ellipses highlight major clusters and black stars mark prototype centroids. This visualization verifies whether adapters exhibit task-separable low-dimensional structure.
• Figure 3: Prototype activation heatmap across disease types. Rows correspond to prototypes and columns to disease categories. Values show normalized mean activation of each prototype for each disease; hierarchical clustering reorders rows and columns to emphasize specialization patterns.
• Figure 4: Performance as a function of support-set size. Curves show AUC versus support set size (5, 10, 20, 50) for multiple disease cohorts. Error bars indicate per-size variability (standard deviation or bootstrap confidence intervals). The plot demonstrates few-shot scaling behaviour.
• Figure 5: Stability of key metrics across random seeds. Boxplots summarize the distribution of AUC, F1 and ECE across repeated runs with different seeds, illustrating model robustness to initialization and data shuffle.