Harvard's COMPASS Model Predicts Immunotherapy Response Across Cancers with Concept Bottleneck Transformer
核心洞察
COMPASS, a concept bottleneck transformer developed at Harvard Medical School, predicts immunotherapy response across seven cancer types and multiple ICI regimens, outperforming 22 existing methods by 8.5% in accuracy and 15.7% in AUPRC.
The model was pretrained on 10,184 tumors from TCGA and fine-tuned on 1,133 patients across 16 clinical cohorts, using 44 interpretable immune concepts rather than a black-box approach.
In the IMvigor210 atezolizumab trial, COMPASS-predicted responders showed a 1-year overall survival rate of 86% versus 40% for non-responders (HR = 4.7, P = 1.7 × 10⁻⁷), outperforming TMB and PD-L1 (搜索) scoring.
A team at Harvard Medical School has developed COMPASS, a concept bottleneck transformer model that predicts immunotherapy response across multiple cancer types and immune checkpoint inhibitor (ICI) regimens, addressing what senior author Marinka Zitnik, associate professor of biomedical informatics at Harvard Medical School, described as "one of the central unsolved problems in oncology." Published this week in Nature Medicine, the model outperformed 22 existing prediction methods and offers interpretable, biology-grounded rationales for its predictions.
The challenge COMPASS tackles is well known: response rates to ICIs range from 10% to 40% depending on cancer type, yet identifying which patients will benefit has remained frustratingly difficult due to small cohort sizes, patient heterogeneity, and complex tumor–immune interactions.
A biologically grounded architecture
COMPASS departs from conventional black-box models by routing approximately 15,672 protein-coding genes through 44 human-readable immune concepts before generating a response prediction. These concepts — curated from 132 gene signatures drawn from the literature — capture immune cell types, functional states, signaling pathways, and non-immune tumor biology such as stromal programs and DNA damage response.
The model uses a transformer-based gene language model to encode expression profiles, then projects gene embeddings onto the 132 gene signatures, which are aggregated into 43 high-level tumor immune microenvironment (TIME) concepts. A cancer type token is included as an additional concept, yielding 44-dimensional patient embeddings.
COMPASS was pretrained on transcriptomes from 10,184 tumors across 33 cancer types in The Cancer Genome Atlas (TCGA) using self-supervised contrastive learning. It was subsequently fine-tuned on 1,133 patients from 16 clinical cohorts spanning seven cancer types and multiple ICI regimens — including anti-PD-1 (搜索), anti-PD-L1 (搜索), anti-CTLA-4 (搜索), and combinations — with cohort sizes ranging from 16 to 298 patients.
State-of-the-art performance across benchmarks
In a leave-one-cohort-out evaluation designed to assess generalizability in clinically relevant scenarios, COMPASS substantially outperformed 22 widely used ICI response prediction methods. Partial fine-tuning (COMPASS-PFT) and linear probing (COMPASS-LFT) improved accuracy by an average of 8.5%, area under the precision-recall curve (AUPRC) by 15.7%, and Matthews correlation coefficient (MCC) by 12.3% compared to the second-best-performing models.
Performance remained robust across large, medium, and small cohorts. In cohort-to-cohort transfer analyses — where models trained on one cohort predicted responses in an entirely different cohort — COMPASS-LFT achieved successful model transfer in 163 of 240 cases, substantially exceeding the top-performing published methods including PGM (130/240), Teff (118/240), and NetBio (117/240).
Cross-indication and cross-therapy generalizability
COMPASS demonstrated the ability to generalize beyond individual studies. In cross-indication prediction, when lung adenocarcinoma (LUAD) was excluded from training, COMPASS-PFT achieved 76.5% accuracy on the held-out cohort. In cross-therapy evaluations, the model achieved 70.8% accuracy for anti-CTLA-4 (搜索) when trained on PD-1 (搜索)/PD-L1 (搜索) cohorts, indicating shared immune mechanisms.
Notably, when predicting responses to combination therapies (ipilimumab plus pembrolizumab) using models trained only on monotherapy cohorts, COMPASS-PFT achieved 85.3% accuracy.
Multi-stage fine-tuning for drug-specific models
COMPASS employs a multi-stage fine-tuning (MSFT) strategy to build robust, treatment-specific models from small clinical cohorts. For atezolizumab response prediction in kidney renal clear cell carcinoma (KIRC, n = 89), MSFT achieved 73.7% accuracy compared to 70.3% for single-stage fine-tuning on the drug-specific cohort alone. For pembrolizumab response in LUAD (n = 33), MSFT reached 91% accuracy versus 67% for single-stage fine-tuning.
Survival prediction and biomarker comparison
Applied to the IMvigor210 phase 2 trial of atezolizumab in metastatic urothelial carcinoma (搜索) (n = 298), COMPASS demonstrated strong prognostic capability. Patients classified as responders (response probability ≥ 0.5) showed a 1-year overall survival rate of 86% compared to 40% for non-responders, yielding a hazard ratio of 4.7 (log-rank P = 1.7 × 10⁻⁷).
In a subset of 234 patients with available TMB data, COMPASS (HR = 4.37, 95% CI: 2.29–8.32) outperformed TMB (HR = 1.67, P = 0.0038), PD-L1 (搜索) IC2+ scoring (HR = 1.75, P = 0.0018), and IHC-based immune phenotype (HR = 1.85, P = 0.0042).
Revealing resistance mechanisms
COMPASS's interpretable concept layer enabled detailed characterization of resistance mechanisms. Inflamed tumors showed high activation of proinflammatory concepts including Cytotoxic T Cell, IFNγ Pathway, Immune Checkpoint, and Macrophage. Desert tumors lacked proinflammatory activation and exhibited elevated scores for dysfunctional immune components including NK Cell, Innate Lymphoid Cell, and B Cell General.
Among inflamed non-responders — patients who would conventionally be predicted to respond — COMPASS identified distinct resistance mechanisms: one cluster showed strong Endothelial concept activation reflecting angiogenesis and vascular remodeling, while another exhibited high TGFβ Pathway activation consistent with stromal remodeling and T cell exclusion. Additional clusters displayed CD4+ T cell immunosuppression and B cell deficiency.
Personalized response maps
COMPASS generates personalized response maps that trace how individual patient gene expression propagates through biological pathways to produce predictions. These maps span five levels: gene expression, encoder representations, granular immune concepts, aggregated TIME concepts, and final response probability. An inflamed responder might show broad IFNγ and cytotoxic activation with minimal immunosuppression (PR = 1.0), while an inflamed non-responder could display co-activation of TGFβ signaling and B cell deficiency (PR = 0.22).
All analyses remain retrospective, using deidentified datasets from published trials and TCGA. The model's performance in prospective clinical settings with real-world sequencing panels remains to be evaluated.