Towards a foundation model for treatment-related adverse events in cancer immunotherapy

Background: Immune-related adverse events (irAEs) provide important clues about immune activation during checkpoint blockade, yet real-world AE reporting is often sparse and fragmented. A further challenge is that FAERS provides unstructured case-level reports, while ClinicalTrials.gov reports arm-level AE tables that differ greatly in completeness and level of detail across trials. These differences make it difficult to study toxicity patterns across data sources in a unified way. To address this gap, we developed a denoising AE foundation model that reconstructs missing Preferred Terms (PTs), learns latent AE co-occurrence structure, and generates biologically interpretable AE embeddings. Methods: We collected ten years of FAERS reports related to immune checkpoint inhibitors (2015Q1-2025Q2) and curated a vocabulary of 200 treatment-related PTs. The model treats each AE profile as partially observed and performs PT-aware masking, case-level PT dropping, and multi-label reconstruction on a BioBERT backbone. Candidate-set validation was used to evaluate recovery of truly unobserved PTs, and mid-layer pooling was applied to improve stability across diverse AE profiles. For ClinicalTrials.gov, we harmonized arm-level AE tables by standardizing differences in reporting depth and AE granularity. We then generated pooled AE signatures for each treatment arm and evaluated whether these signatures could distinguish immunotherapy arms with known differences in clinical outcomes. Results: The model recovered intentionally hidden AEs with strong performance (Recall@5 = 0.38; Recall@10 = 0.51). The learned embedding space separated well-known irAEs from non-immune toxicities and formed clusters that aligned with MedDRA SOC patterns. When applied to ClinicalTrials.gov treatment arms, the AE signatures reflected cancer-type-specific patterns, including dermatologic events in melanoma and hepatic or gastrointestinal patterns in hepatocellular carcinoma and GI cancers. Arm-level embedding profiles also differentiated treatment arms with superior outcomes across several ICI trials, suggesting that the representation space captures meaningful biology related to treatment response. Conclusions: We present a foundation model that integrates FAERS and ClinicalTrials.gov AE data to denoise AE profiles, learn biologically grounded toxicity embeddings, and reveal AE patterns associated with clinical benefit. This framework provides a scalable approach for studying AE biology and may support AE-based biomarker development and arm-level prediction of immunotherapy efficacy.