Three interconnected programs spanning melanoma prognosis, immunotherapy toxicity, and the biology of the tumor microenvironment — all grounded in real clinical data from multi-institutional and population-level cohorts.
Most melanoma deaths occur in patients initially diagnosed with early-stage (Stage I–II) disease who later experience recurrence. Our machine learning platform — trained on 1,720 melanomas across MGB and DFCI — predicts which patients face the highest recurrence risk.
Our ML models integrate digital pathology, imaging, and clinical data using advanced computer vision techniques, achieving AUROC 0.845 internally and 0.812 externally — outperforming AJCC staging in predicting recurrence of non-metastatic melanoma. These models are currently being scaled across multiple external institutions.
Permutation importance analysis across Gradient Boosting (GB) and Random Forest (RF) models consistently identified Breslow thickness and mitotic rate as the two dominant predictors — validating established pathologic risk factors while also surfacing novel signals such as TIL type and radial growth pattern.
The model also generates patient-level time-to-event recurrence trajectories, enabling individualized surveillance planning and counseling for patients with Stage I–II melanoma.
Our ongoing work focuses on integrating foundation models with attention-based multiple instance learning on whole-slide images to extract prognostic signal invisible to the human eye.
Our seminal population-level analysis of 8,637 ICI recipients and 8,637 matched controls established the true epidemiology of cutaneous immune-related adverse events (irAEs) — the spectrum of skin toxicities from cancer immunotherapy, including rash, pruritus (itch), inflammatory skin conditions, vitiligo, and lichenoid eruptions — finding that only 10 of 43 previously reported dermatoses are actually elevated in ICI-treated patients.
In our landmark Lancet Oncology 2024 study, we characterized the downstream implications of multi-organ irAEs across 12 organ systems in large MGB and TriNetX cohorts. Using non-negative matrix factorization, we identified seven distinct irAE clusters — finding that endocrine- and cutaneous-predominant clusters are associated with improved survival, while respiratory- and neurologic-predominant clusters confer worse outcomes.
Additional work across 14,016 patients demonstrated that cutaneous irAEs are associated with a 22% reduction in mortality, with specific morphologies (vitiligo, lichenoid, acneiform) predicting significantly improved survival — positioning irAEs as actionable biomarkers of immunotherapy response.
Our group has built one of the largest multi-institutional irAE registries and biorepositories, encompassing 24,000+ immunotherapy recipients treated between 2011 and 2025, leveraging AI and computational phenotyping to characterize toxicity patterns and treatment outcomes at scale.
In collaboration with the Broad Institute and Dana-Farber, we conducted the first large-scale GWAS of immune checkpoint inhibitor toxicity — identifying 3 genome-wide significant loci, including a variant near IL7 (p = 3.6×10⁻¹¹, HR 2.1) that colocalized with a cryptic exon affecting lymphocyte homeostasis.
Patients carrying this variant showed increased lymphocyte stability after ICI initiation, which was itself predictive of downstream irAEs and associated with improved survival — uncovering a mechanistic link between host autoimmunity and immunotherapy response.
Nature Medicine 2022
Our spatial biology program uses multiplex immunofluorescence imaging, spatial transcriptomics, and multiomic profiling of melanoma tissue to map cell-cell interactions within the tumor microenvironment — with a focus on in-transit melanoma as a model for studying stromal remodeling and immune evasion.
By integrating single-cell and spatial resolution data, we identify biomarkers of treatment response, mechanisms of immune checkpoint blockade resistance, and spatially resolved signatures of melanoma progression.
We developed SpatialCells, an open-source software package for automated, region-based profiling of tumor microenvironments using spatially resolved multiplexed single-cell data — enabling high-throughput feature extraction and machine learning predictions from complex tissue specimens.
Wan et al., Briefings in Bioinformatics 2024
