Accurate and scalable multi-disease classification from adaptive immune repertoires

Abstract

BackgroundMachine learning models trained on paratope-similarity networks have shown superior accuracy compared with clonotype-based models in binary disease classification. However, the computational demands of paratope networks hinder their use on large datasets and multi-disease classification. MethodsWe reanalyzed publicly available T cell receptor (TCR) repertoire data from 1,421 donors across 15 disease groups and a large control group, encompassing approximately 81 million TCR sequences. To address computational bottlenecks, we replaced the paratope-similarity network approach (Paratope Cluster Occupancy or PCO) with a new Fast Approximate Clustering Techniques (FACTS) pipeline, which is comprised of four main steps: (1) high-dimensional vector encoding of sequences; (2) efficient clustering of the resulting vectors; (3) donor-level feature construction from cluster distributions; and (4) gradient-boosted decision tree classification for multi-class disease prediction. FindingsFACTS processed 107 sequences in under 120 CPU hours. Using only TCR data, and evaluated with 5-fold cross-validation, it achieved a mean ROC AUC of 0.99 across 16 disease classes. Compared with the recently reported Mal-ID model, FACTS achieved higher donor-level classification accuracy for BCR (0.840 vs. 0.740), TCR (0.882 vs. 0.751), and combined BCR+TCR datasets (0.904 vs. 0.853) on the six-class Mal-ID benchmark. FACTS also preserved biologically meaningful signals, as shown by unsupervised t-SNE projections revealing distinct disease-associated and potentially age-associated clusters. InterpretationParatope-based encoding with FACTS-derived features provides a scalable and biologically grounded approach for adaptive immune receptor (AIR) repertoire classification. The resulting classifier achieves superior multi-disease diagnostic performance while maintaining interpretability, supporting its potential for clinical and population-scale health profiling. FundingThis study was supported by the Japan Society for the Promotion of Science (JSPS) KAKENHI [JA23H034980], the Japan Agency for Medical Research and Development (AMED) [JP25am0101001], and the Kishimoto Foundation Fellowship. Research in contextO_ST_ABSEvidence before this studyC_ST_ABST and B cell receptor (TCR and BCR) repertoires encode lifelong immunological memory and antigen-specific responses, making them valuable biomarkers for disease diagnosis and prediction. Existing machine learning (ML) models for adaptive immune receptor (AIR) repertoires often rely on clonotype-based representations, which limit shared receptor detection between donors and thus reduce cross-individual disease signature detection. Most models also lack robust multi-disease, population-scale performance. Our previous work showed that representing repertoires as paratope-similarity networks increased the fraction of shared receptors between donors and improved disease classification. However, their computational complexity has limited their scalability for the large datasets required in multi-disease classification. Added value of this studyWe introduce FACTS, a unified ML framework integrating paratope similarity with scalable sequence encoding. Applied to TCR repertoires from 1,421 donors across 15 diseases and one control group, FACTS maintained high performance while efficiently processing 81 million sequences on standard CPU infrastructure. Compared to Mal-ID, our paratope-encoded method achieved significantly higher donor-level accuracy and revealed biologically meaningful disease- and potentially age-associated patterns. Implications of all the available evidenceFACTS offers high accuracy, and interpretability for multi-disease classification, bringing AIR repertoire-based diagnostics closer to clinical translation and potentially guiding precision immunotherapy and immune-based therapeutic discovery for a wide range of diseases.