Background: Rheumatoid arthritis (RA) constitutes a pervasive clinical challenge characterized by profound phenotypic heterogeneity, intricate comorbidity profiles—particularly with depression—and highly variable responses to biological therapies. Conventional predictive frameworks frequently struggle to synthesize these multidimensional complexities, primarily due to their inability to resolve the “curse of dimensionality” inherent in sparse biomedical data.

Objectives: To develop and validate a unified foundation-like framework, the TabPFN-based Rheumatoid Arthritis Intelligence Network (TRAIN), capable of integrating million-scale multi-omics data to overcome data sparsity. The study aims to systematically evaluate TRAIN’s performance in three critical clinical tasks: precise disease diagnosis, stratification of depression comorbidity, and forecasting therapeutic response to biological agents.

Methods: By harmonizing a massive data foundation comprising approximately 3 million integrated data points derived from 103,099 longitudinal real-world clinical records and 5,619 high-quality public multi-omics profiles, TRAIN effectively surmounts the limitations of traditional small-sample training. Mechanistically, the framework employs a robust self-supervised manifold learning strategy to extract universal latent representations from unlabeled data, thereby capturing the intrinsic biological syntax governing disease progression without relying on simplistic linear assumptions. We systematically validated the predictive fidelity of TRAIN across three pivotal clinical dimensions.

Results: In the diagnostic task distinguishing RA from healthy controls, the synergistic fusion of transcriptomic and proteomic signatures with clinical phenotypes eliminated diagnostic ambiguity, achieving near-perfect discrimination with an area under the curve (AUC) of 0.997 ± 0.008. For the stratification of comorbidities, TRAIN successfully decoupled the latent biological signals of depression from systemic inflammation, achieving an AUC of 0.959 ± 0.055 and resolving a long-standing challenge in distinguishing somatic symptoms from psychiatric overlap. Furthermore, in the critical task of forecasting therapeutic response, the model demonstrated remarkable generalization capabilities, accurately identifying non-responders even within data-constrained prognostic scenarios with a leading AUC of 0.972 ± 0.062. Crucially, our multi-omics fusion approach confirms that integrating heterogeneous data streams yields a predictive stability that significantly outperforms representative state-of-the-art machine learning architectures.

Conclusions: By converting the challenge of high-dimensional data sparsity into an advantage of knowledge transfer, this work establishes TRAIN as a definitive methodological advancement, providing a stable, precise, and unified computational foundation for personalized RA diagnosis and therapeutic decision-making that bridges the gap between complex multimodal data and precision medicine.

REFERENCES: NIL.

Acknowledgments: NIL.

Disclosure of Interests: None declared.