Path2Marker: Cell-level prediction of multiplex protein expression from routine H&E slides

Background: High-plex spatial proteomics platforms (e.g., PhenoCycler) have transformed our ability to map the tumor microenvironment (TME) at single-cell resolution, but their cost constrains cohort sizes for biomarker discovery and validation. Recently, ROSIE (Wu et al., Nat Commun 2025) aimed to tackle this problem by predicting protein markers directly from H&E stained slides. However, the number of robustly predicted markers obtained by it has been fairly limited, with only 5 markers reaching a Pearson r correlation of above 0.4 between measured and predicted marker intensity. Here, we present Path2Marker, a cell-level deep learning framework that substantially expands the panel of robustly predicted protein markers in multiple tumor types. Methods: We analyzed three new cancer-specific PhenoCycler (CODEX) cohorts with paired H&E and multiplex immunofluorescence, including lung cancer (88 samples, 660791 cells), colorectal cancer (106 samples, 624919 cells) and breast cancer (115 samples, 901684 cells), each stained with a 55 proteomic marker panel. For each disease, we trained a cancer-specific model to predict per-cell marker intensities from the H&E slides, and additionally evaluated an ensemble model that averages predictions from all three cancer-specific models. Model performance was evaluated on 60 held-out samples (all three diseases; 396,995 cells), using Pearson correlation between the measured and predicted marker intensities. Results : We robustly predict (Pearson r > 0.4 for measured vs. predicted intensity) 23, 26, and 44 markers in the breast, lung and colorectal cohorts, respectively, markedly outperforming the published state of the art. When benchmarked on the same samples, ROSIE achieved fewer robustly predicted markers, with only 3 markers in colon, 2 in lung and 0 in breast. The ensemble model significantly improved mean marker-level correlation in lung and was comparable to the indication-specific models in breast and colorectal cancer. Top-performing markers included EpCAM in colon (r=0.79), PanCK in lung (r=0.73), and PanCK in breast (r=0.64). Notably, they include not only lineage markers (e.g., PanCK, CD3e) but also functional markers (e.g., PD-L1, Ki-67), enabling downstream cell-state and cell-type annotation, laying the basis for robust annotation of 25, 16, and 17 different cell-types in colon, lung, and breast, respectively. Conclusion : Path2Marker enables robust prediction of more than 20 multiplex protein markers at single-cell, spatial resolution directly from standard H&E slides across three major solid tumors. It markedly improves upon the predictive accuracy of extant tools, opening up the possibility of fast and low-cost annotation of large cancer patients cohorts directly from the histopathology slides.