Deciphering dermatological distinctions: Cornulin as a discriminant biomarker between basal cell carcinoma and squamous cell carcinoma detected through e‐biopsy and machine learning

• Published in: Experimental dermatology 2024-06, Vol.33 (6), p.n/a
• Main Authors: Vitkin, Edward, Wise, Julia, Berl, Ariel, Shir‐az, Ofir, Kravtsov, Vladimir, Yakhini, Zohar, Shalom, Avshalom, Golberg, Alexander
• Format: Article
• Language: English
• Subjects: Basal cell carcinoma; Biomarkers; Biopsy; Cancer; cornulin; diagnostics; Electroporation; e‐biopsy; Immunohistochemistry; Learning algorithms; Machine learning; Metastases; molecular pathology; Morbidity; Patients; Proteins; protein–protein interaction; Proteomes; Sampling; skin cancer; Squamous cell carcinoma; Transcriptomics
• ISSN: 0906-6705; 1600-0625
• DOI: 10.1111/exd.15124

Clinical misdiagnosis between cutaneous squamous cell carcinoma (cSCC) and basal cell carcinoma (BCC) poses treatment challenges and carries risks of recurrence, metastases and increased morbidity and mortality. We aimed to identify discriminant proteins markers for cSCC and BCC using a minimally invasive proteome sampling method called e‐biopsy, employing electroporation for non‐thermal cell permeabilization and machine learning. E‐biopsy facilitated ex vivo proteome extraction from 21 cSCC and 21 BCC pathologically validated human cancers. LC/MS/MS profiling of 126 proteomes was followed by machine learning analysis to identify proteins distinguishing cSCC from BCC. For identified panel validation, we used proteomes sampled by e‐biopsy from unrelated 20 cSCC and 46 BCC human cancers, and differential expression analysis of published transcriptomics. Cornulin, the most commonly chosen discriminative biomarker by machine learning models, was also validated using fluorescent immunohistochemistry. One hundred and ninety‐two proteomes sampled from one hundred eight patients were analysed. Machine learning‐based approaches resulted in a set of 11 potential biomarker proteins that can be used to construct a patient classification model with 95.2% average cross‐validation accuracy, BCC precision of 93.6 ± 14.5%, cSCC precision of 98.4 ± 7.2%, specificity of 97.7 ± 11.8% and sensitivity 92.7 ± 15.3%. Protein–protein interaction analysis revealed a novel interaction network connecting 10 of the 11 resulted proteins. Histological and transcriptomic validation confirmed cornulin as a discriminant marker significantly lower in cSCC than in BCC. E‐biopsy combined with machine learning provides a novel approach to molecular sampling from skin for biomarker detection and differential expression analysis between cSCC and BCC.