Can the Human Plastisphere Account for the Rise of Type 1 Diabetes: A Computational Molecular Mimicry Study

The HLA molecules and respective epitopes (between parenthesis) are HLA-A*02:01 (RLLPLLALLAL), HLA-B*08:01 (TPKTRREAEDL); HLA-DQA1*03:01 (FVNQHLCGSHLVEAL); HLA-DQA1*05:01 (SHLVEALYLVCGERG), HLA-DQB1*02:01 (SHLVEALYLVCGERG), HLA-DQB1*03:02 (SHLVEALYLVCGERG), HLA-DRB1*03:01 (LPKPPKPVSKMRMATPLLMQALPM), HLA DRB1*04:01 (NFFRMVISNPAAT), HLA-DRB3*02:02 (SPLGQSQPTVAGQPSARPAAEEYGYIVTDQKPLSLAAGVK), and HLA-DRB4*01:01 (KVNFFRMVISNPAATHQD) were used as archetypes of T1DM antigen epitopes that bind HLA molecules associated with increasing risk of early onset of T1DM. The amino acid sequences of the HLA molecules were obtained from (https://www.ebi.ac.uk/ipd/imgt/hla/ ) and depicted in Table S1. The corresponding antigen epitopes were obtained from IEDB (https://www.iedb.org/) and entered in Table S2. Fragments of six common microplastics, high-density polyethylene (HDPE), low-density polyethylene (LDPE), polyethylene terephthalate (PET), polypropylene (PP), and polystyrene (PS) and polyvinyl chloride (PVC) were chosen as potential candidates for molecular mimicry. Their SMILES fragments were retrieved from PubChem (https://pubchem.ncbi.nlm.nih.gov/) and were input into a python v.3.11.9 (with the RdKit library and AllChem module) script to generate covalently linked microplastic oligomers of desired length with corresponding structural data files (SDF). These SDF files were input into UCSF’s ChimeraX (v.1.10.1) for structure analysis. The polymer length was measured using the ChimeraX ‘distance’ command. All six of the plastics fragments were selected to be of the approximate length of the binding region of the target Human Leukocyte Antigen (HLA) molecules. The SMILE notation for all the plastic fragments used here are entered into Figure S1. Structural modeling of the HLA-T1DM epitope complexes and HLA-plastic fragments complexes were done using Boltz-2 (v.2.2.0), an open-source AI affinity prediction model. Boltz‑2 was executed on the Granite GPU cluster at the Utah Center for High Performance Computing (https://www.chpc.utah.edu/). Boltz-2 was run using standard parameters by enabling the MSA server, specifying 10 recycling steps, generating 5 diffusion samples, and performing 500 sampling steps. Each run generated 4 candidate models for each FASTA file (in CIF file format), but model 0 consistently exhibited the highest accuracy and highest confidence scores. The output CIF files for each structural model were input into ChimeraX to visualize the calculated structure and to calculate structural similarity. All the Boltz-2 results are available at 10.5281/zenodo.19100817. The location of the T1DM epitopes and microplastic polymers were visually inspected to determine whether they resided in the binding region of HLA molecule and depicted in Figure S2. The biding facts discussed under the results section of the paper have been extracted from the .cif files for all the models discussed using the University of Utah instance of ChatGPT 5.2 on March 11, 2026. All results reported here were furthermore verified by one of the authors (JCF) and none of these results we directly imported from ChatGPT without human supervision as in many cases the initial ChatGPT results we clearly erroneous and needed further prompting to clarify them. Clearly LLMs like ChatGPT can be used to more efficiently analyze .cif files, but the technology is not currently accurate to allow automatic analysis without expert supervision.