BPS2026 – Biophysics-based featurization as a mean for deep learning applications on early-stage diagnosis of autoimmune disorders with ambiguous clinical data

Autoimmune disorders occur when the immune system mistakenly attacks the body’s own tissues. This loss of immune tolerance and the development of autoimmunity are influenced by human leukocyte antigen (HLA) molecules. A well-established way to statistically connect an HLA allele to an autoimmune disorder is to perform association tests. Yet, the number of tests needed can scale combinatorically with the number of alleles probed, and results found for one allele often cannot be extrapolated to alleles not included in the tests. We approach these challenges by characterizing HLA alleles by their biophysical features extracted from our in-house deep learning model, HLA inception. Our model was first trained on HLA alleles’ electrostatic maps, followed by post processing to predict allele-peptide binding properties. These trained, electrostatics-centric embeddings were then used as features characterizing HLA alleles. Unsupervised clustering of alleles in this feature space correctly grouped alleles according to their autoimmune disorder associations and successfully distinguished autoimmune disorders from non-autoimmune diseases. Our results also demonstrated a noticeably higher signal-to-noise ratio when being contrasted to baseline clustering models utilizing natural language processing (NLP) without any biophysics nor bioinformatics-based feature. Interestingly, our results remain robust even when clinical data with ambiguous measurements ware included. Our approach enables the estimation of an allele’s connection to autoimmunity through its biophysical features at a molecular level without prior association tests, providing a pathway to diagnose patients with alleles accompanied by limited clinical data. It also allows us to construct an autoimmunity-associated peptide library, to be validated by our Mayo Clinic collaborators. Lastly, we discuss implementable steps to examine the effectiveness of our approach versus more popular embedding methods, including ESM embeddings and BRET. HLA inception website: https://www.strsysbiolab.academy/software/hla.