TPC Data Enables AI Model for Predicting Protein Expression from Standard Pathology Images

Research published in Nature Machine Intelligence demonstrates that multimodal data from the Tumor Profiler Center can train artificial intelligence models to predict spatially resolved protein expression patterns directly from routine histopathology slides. The study, published on August 4, 2025, presents HistoPlexer, a deep learning framework with potential to make advanced molecular profiling more accessible in precision oncology.

Addressing a Key Bottleneck in Precision Oncology

Understanding the tumor microenvironment—the complex ecosystem of tumor cells, immune cells, and supporting tissue—is critical for predicting treatment response and guiding therapy selection. Multiplexed protein imaging technologies such as imaging mass cytometry (IMC) can simultaneously measure dozens of protein markers at single-cell resolution, revealing how different cell types interact within tumors. However, these technologies require specialized equipment, are costly and time-intensive, and can only analyze small tissue regions, limiting their clinical adoption.

In contrast, haematoxylin and eosin (H&E) staining remains the universal standard in pathology. Every cancer patient's tissue is routinely stained with H&E for diagnosis, making these slides abundant and immediately available. The key question addressed by this research is whether the morphological information visible in H&E images contains sufficient information to computationally predict protein expression patterns that would otherwise require expensive multiplexed imaging.

The HistoPlexer Approach

The research team developed HistoPlexer using data from the TPC's TuPro study, which provided paired H&E and IMC images from 336 tissue regions across 78 metastatic melanoma patients. This multimodal dataset—where the same tumor tissue was analyzed with both standard pathology and multiplexed protein imaging—was essential for training the model to learn the relationship between tissue morphology and protein expression.

HistoPlexer employs a conditional generative adversarial network architecture that simultaneously predicts 11 protein markers from H&E images. These markers include tumor cell markers (MelanA, S100, gp100, SOX10), immune cell markers (CD3, CD8a, CD20, CD16, CD31), and antigen presentation markers (HLA-ABC, HLA-DR). Critically, the model predicts all markers simultaneously rather than independently, which the authors demonstrate better preserves biologically meaningful spatial relationships between different proteins.

Validation and Clinical Relevance

The study includes rigorous validation across multiple dimensions. Expert pathologists evaluated the generated protein maps and found them realistic, with the majority indistinguishable from actual IMC images. Importantly, the model preserves spatial co-localization patterns—when two proteins are expressed together or separately in real tissue, the predicted maps maintain these relationships.

For clinical utility, the authors demonstrate that HistoPlexer can stratify patients into immune phenotypes based on CD8+ T cell infiltration patterns. This classification into "immune-hot" (high infiltration) versus "immune-cold" (low infiltration) tumors is clinically relevant for immunotherapy treatment decisions. When tested on an independent cohort of 472 melanoma patients from The Cancer Genome Atlas, adding HistoPlexer-generated features to predictive models improved survival prediction by 3.18% and immune subtype classification by 17.02% compared to using H&E images alone.

Implications for Precision Medicine

By generating protein expression maps from standard histopathology slides, HistoPlexer offers a path to democratize access to molecular profiling. While the model does not replace actual multiplexed imaging for research purposes, it could enable tumor microenvironment characterization at institutions without access to specialized imaging equipment. This is particularly relevant for patient stratification, where understanding immune infiltration patterns can inform treatment decisions without requiring the detailed single-cell resolution needed for mechanistic studies.

The study exemplifies how large-scale, high-quality multimodal datasets—such as those generated by the TPC consortium—enable advances in computational pathology and artificial intelligence for medicine. The paired H&E and IMC data from the TuPro study provided the essential training foundation for this work, demonstrating the value of systematic multimodal tumor profiling efforts.