Generalizable AI predicts immunotherapy outcomes across cancers and treatments

The paper introduces COMPASS, a pan-cancer foundation model that leverages a concept-bottleneck transformer to accurately predict immunotherapy response across diverse cancers and treatments by encoding gene expression into 44 biologically grounded immune concepts, thereby outperforming existing baselines while providing interpretable mechanistic insights into resistance.

The Gist

Imagine you are a doctor trying to decide which cancer patients should receive a powerful new treatment called immunotherapy. This treatment is like a "supercharger" for the patient's own immune system, teaching it to recognize and destroy cancer cells.

The problem? It's a hit-or-miss game. Some patients get amazing results, while others see no benefit at all. Currently, doctors use a few "clues" (biomarkers) to guess who will respond, but these clues are often like looking at a car through a foggy windshield—they give a vague idea, but not a clear picture.

Enter COMPASS, a new AI tool described in this paper. Think of COMPASS not just as a calculator, but as a master translator and a detective.

The Problem: The "Black Box" of Cancer

Cancer is incredibly complex. It's like a chaotic city where the immune system (the police) is trying to fight the cancer (the criminals). Sometimes the police are overwhelmed, sometimes they are confused, and sometimes they are just tired.

Old methods tried to predict the outcome by counting specific things, like "How many police cars are there?" (T-cell count) or "How many criminals are there?" (Tumor Mutational Burden). But this is like judging a whole war just by counting the number of tanks. It misses the bigger picture: Are the police tired? Are the roads blocked? Is the police chief giving bad orders?

The Solution: COMPASS, the "Concept Translator"

The researchers built COMPASS using a clever trick called a Concept Bottleneck.

Imagine you are trying to explain a complex movie plot to a friend who speaks a different language.

• Old AI models try to translate every single word of the script directly into the other language. This is messy and prone to errors.
• COMPASS works differently. It first translates the script into 44 key "concepts" that anyone can understand, like "The Hero is brave," "The Villain is hiding," or "The weather is stormy." Only after understanding these concepts does it predict the ending.

In the world of cancer, these 44 concepts are biological ideas like:

• "Are the T-cells exhausted?" (Like a tired police officer).
• "Is the TGF-beta pathway active?" (Like a roadblock preventing police from entering the city).
• "Is there a B-cell deficiency?" (Like a lack of backup units).

How It Learned: The "Gym" and the "Specialist"

COMPASS didn't start out knowing everything. It went through two stages of training:

1. The Gym (Pre-training): First, the AI looked at data from 10,000+ patients across 33 different types of cancer. It didn't know who would respond to treatment yet; it just learned the "language" of cancer. It learned what a "tired immune system" looks like in a lung tumor versus a skin tumor. It's like a student reading every medical textbook in the library to understand the basics of biology.
2. The Specialist (Fine-tuning): Then, the AI was shown data from 16 specific clinical trials where patients actually received immunotherapy. It learned to connect those 44 concepts to real-world results: "Ah, when the 'TGF-beta roadblock' concept is high, the patient usually doesn't respond."

Why It's a Game-Changer

The paper shows that COMPASS is significantly better than existing methods. Here is why, using simple analogies:

• It's a Universal Translator: Old models were like specialists who only spoke "Lung Cancer." If you showed them "Kidney Cancer," they were confused. COMPASS learned the universal language of the immune system, so it can predict outcomes for new cancer types and new drugs it has never seen before.
• It Sees the Invisible: Sometimes a patient looks like they should respond (the "police" are present), but they don't. COMPASS found out why. It discovered that in these "fake responders," the police were being blocked by a "TGF-beta roadblock" or were "exhausted." It explains the mechanism of failure, not just the failure itself.
• It Saves Lives: In a test with bladder cancer patients, COMPASS correctly identified who would live longer. It was much more accurate than the current standard tests (like checking for PD-L1 or counting mutations).

The "Personalized Map"

One of the coolest features is the Personalized Response Map.
Imagine you get a report card. Instead of just a grade (Pass/Fail), COMPASS gives you a map.

• "Your tumor has high 'Cytotoxic T-cell' activity (Good!)"
• "But it also has high 'Endothelial Exclusion' (Bad! This is blocking the cells)."
• "Therefore, the prediction is: Low chance of response."

This map helps doctors understand why a patient might not respond and could even suggest new combinations of drugs to break those roadblocks.

The Bottom Line

COMPASS is like a highly trained, biologically grounded detective. Instead of just guessing based on a few clues, it understands the entire story of the tumor's interaction with the immune system. It translates complex genetic data into 44 understandable "concepts," allowing doctors to predict who will benefit from immunotherapy with much higher accuracy and to understand the biological reasons behind the prediction.

While it's not ready to replace doctors in the clinic tomorrow (it still needs more testing), it represents a massive leap forward in making cancer treatment more precise, personalized, and effective.

Technical Summary

1. Problem Statement

Immune checkpoint inhibitors (ICIs) have revolutionized cancer treatment, but clinical benefit remains uneven. Only a minority of patients achieve durable responses, and existing biomarkers (e.g., Tumor Mutational Burden [TMB], PD-L1 expression) suffer from poor generalizability across different tumor types, drugs, and clinical settings.

• Limitations of Current Methods: Transcriptomic signatures (e.g., TIDE, IMPRES) and machine learning models often rely on fixed gene interactions or are specific to single cancer types. They fail to capture the complex, heterogeneous biological mechanisms of resistance, particularly in "immune-inflamed" patients who nonetheless fail to respond to therapy.
• The Gap: There is a need for a pan-cancer foundation model that can predict immunotherapy response with high accuracy, generalize to unseen cancer types and treatments, and provide mechanistic interpretability to explain why a patient is predicted to respond or resist.

2. Methodology: The COMPASS Model

The authors developed COMPASS (Concept-bottleneck Omni-cancer Pan-cancer AI for Survival and Stratification), a foundation model designed to predict ICI response from bulk tumor transcriptomes.

A. Architecture: Concept-Bottleneck Transformer

Unlike standard "black-box" deep learning models, COMPASS uses a Concept-Bottleneck architecture. This forces the model to route gene expression data through human-interpretable biological concepts before making a prediction.

1. Gene Language Model (Encoder): A transformer-based encoder processes the expression profiles of 15,672 protein-coding genes. It uses a learnable gene-specific positional bias (rather than fixed positional encodings) to capture contextual gene-gene interactions.
2. Hierarchical Projector (The Bottleneck):
   • Granular Level: Gene embeddings are projected onto 132 curated gene signatures (representing specific cell types, pathways, and functional states).
   • High-Level Level: These signatures are aggregated into 44 biologically grounded "TIME concepts" (Tumor Immune Microenvironment). These concepts cover immune cell states (e.g., Cytotoxic T cells, B cells), interactions (e.g., Endothelial exclusion), and signaling pathways (e.g., TGF-β, IFN-γ).
   • Cancer Type Token: A specific token is included to account for pan-cancer heterogeneity.
3. Classifier: A lightweight classifier takes the 44-dimensional concept vector as input to predict the probability of response (Responder vs. Non-responder).

B. Training Strategy

The model employs a two-stage transfer learning approach to handle data scarcity in clinical trials:

1. Self-Supervised Pre-training:
   • Data: 10,184 tumor samples from 33 cancer types in The Cancer Genome Atlas (TCGA).
   • Method: Contrastive learning (Triplet loss). The model learns to map augmented views of the same tumor closer together and different tumors further apart in the 44-dimensional concept space. This builds a robust, generalizable representation of TIME without needing outcome labels.
2. Supervised Fine-tuning:
   • Data: 1,133 patients from 16 independent clinical cohorts (7 cancer types, various ICIs).
   • Parameter-Efficient Adaptation: To prevent overfitting on small clinical datasets, COMPASS supports four fine-tuning modes:
     • Full Fine-tuning (FFT): Updates all parameters (for large cohorts).
     • Partial Fine-tuning (PFT): Updates only the projector and classifier (optimal balance).
     • Linear Probing (LFT): Updates only the classifier head (frozen concepts).
     • No Fine-tuning (NFT): Zero-shot prediction using cosine similarity to prototypes in the pre-trained space.

C. Multi-Stage Fine-Tuning (MSFT)

For specific drug or disease development, the authors propose an MSFT strategy: Pre-train $\to$ Fine-tune on pan-cancer ICI data $\to$ Refine on the specific target drug/disease cohort. This transfers shared immune-response features to specific settings.

3. Key Contributions

• First Pan-Cancer Foundation Model for ICI: COMPASS is the first model to successfully predict immunotherapy response across diverse cancer types and drug targets using a unified architecture.
• Interpretability via Concept Bottlenecks: By forcing predictions through 44 biologically defined concepts, the model provides mechanistic insights (e.g., identifying TGF-β signaling or B-cell deficiency as drivers of resistance) rather than just a probability score.
• Robust Generalization: The model demonstrates superior performance in "Leave-One-Cohort-Out" and "Cohort-to-Cohort" transfer scenarios, outperforming 22 existing baseline methods.
• Personalized Response Maps: The system generates visual maps tracing how specific gene activations propagate through biological pathways to the final prediction, aiding hypothesis generation for individual patients.

4. Key Results

• Predictive Performance:
  • In leave-one-cohort-out evaluations, COMPASS (specifically PFT and LFT variants) outperformed 22 baselines (including TIDE, NetBio, and PGM).
  • Metrics: Improved accuracy by 8.5% and Area Under the Precision-Recall Curve (AUPRC) by 15.7% compared to the second-best method.
  • Generalization: Successfully generalized to unseen cancer types (e.g., predicting lung cancer response when trained on others) and unseen drug targets (e.g., predicting anti-CTLA4 response when trained on anti-PD1).
• Survival Analysis:
  • In the held-out IMvigor210 (metastatic urothelial cancer) cohort, COMPASS-stratified responders had significantly longer overall survival compared to non-responders.
  • Hazard Ratio (HR): 4.7 ($p < 0.0001$), significantly outperforming TMB (HR=1.67), PD-L1 (HR=1.75), and IHC phenotypes (HR=1.85).
• Mechanistic Insights:
  • Inflamed Non-Responders: COMPASS identified that patients with "inflamed" phenotypes who failed treatment often exhibited TGF-β signaling, endothelial exclusion, CD4+ T cell dysfunction, and B cell deficiency.
  • Non-Inflamed Responders: Identified residual cytotoxic activity or TMB-associated pathways even in the absence of classic inflammation.
• Multi-Stage Fine-Tuning: MSFT achieved 73.7% accuracy in predicting atezolizumab response in kidney cancer, outperforming single-stage fine-tuning (70.3%) and pan-cancer only models (60.7%).

5. Significance and Impact

• Clinical Trial Optimization: COMPASS can support indication selection and patient stratification in early-phase clinical trials, potentially reducing trial failure rates by identifying the right patients for specific drugs.
• Mechanistic Discovery: The model moves beyond correlation to reveal mechanisms of resistance (e.g., specific signaling pathways blocking T-cell infiltration), offering new targets for combination therapies.
• Translational Readiness: The ability to generalize across platforms, biopsy sites, and cancer types makes COMPASS a practical tool for real-world clinical deployment, provided assay validation is performed.
• Paradigm Shift: The paper demonstrates that combining foundation models (large-scale pre-training) with concept bottlenecks (biological interpretability) is a viable and superior strategy for precision oncology compared to traditional black-box ML or static biomarker scoring.

Limitations Noted: The model relies on bulk RNA-seq (lacking spatial resolution), and the mechanistic concepts require experimental validation. The authors emphasize that COMPASS is currently an exploratory tool for hypothesis generation and trial design, not a standalone clinical decision-making tool.