7-UP

A clinical immunofluorescence workstation images between one and seven protein biomarkers on a tissue section — enough to separate five to seven broad cell types in a tumor. CODEX (co-detection by indexing, commercialized as the Akoya PhenoCycler) pushes that to 40 or more markers on the same slide and resolves the tumor microenvironment in far richer detail, but is slow and costly enough that it rarely reaches routine practice. 7-UP asks whether the missing markers were ever entirely missing: if CD20 reliably co-expresses with CD21 in B cells, and if a proliferating tumor cell advertises its Ki67 status in the shape of its nucleus, then seven measured stains plus each cell's morphology and neighborhood may already encode most of what the other channels would have shown.

Developed by Enable Medicine with collaborators at Stanford University, Columbia University, the University of Pittsburgh, Dana-Farber Cancer Institute, and CellSight Technologies, 7-UP imputes those channels directly. From a 7-plex input panel it predicts single-cell expression for the remaining 33 biomarkers of a 40-plex CODEX panel, and the imputed values are substituted for measured ones in the analyses CODEX is actually run for: cell-type assignment and patient-outcome prediction.

7-UP is the earliest of Enable Medicine's in-silico staining models, preceding ROSIE and Eva, and differs from both in its input: it starts from multiplexed immunofluorescence rather than H&E histology. A model trained on head and neck cancer at one hospital was applied unchanged to two Stanford cohorts and one Dana-Farber cohort — one of them colorectal rather than head and neck cancer — without refitting.

#Key Features

• In silico CODEX: A 7-plex acquisition yields a 40-plex profile per cell, giving laboratories markers they never stained for and freeing panel slots for other antibodies.
• Morphology as evidence: A convolutional featurizer reads each cell's segmentation mask and two progressively wider neighborhood crops, recovering expression that mean intensity misses — vessel-cell F1 rises from 0.44 to 0.67 once morphology is included.
• Cell typing past the panel ceiling: Imputed markers resolve 16 cell types, against the five to seven a 7-plex panel supports on its own.
• Outcome prediction on imputed input: Predicted cell types fed to a graph neural network give HPV status, survival, and recurrence predictions comparable to those from CODEX-measured labels.
• Automatic panel selection: A concrete autoencoder picks the seven most reconstructive markers — DAPI, CD45RA, CD15, pan-cytokeratin, HLA-DR, Ki67, and vimentin — and the step is skipped entirely when the input panel is already fixed.

#Technical Details

Cells are segmented with DeepCell on the DAPI channel and their expressions quantile normalized, arcsinh transformed, and z-scored. For each cell, three image channels are generated per input biomarker — the segmented cell alone at 3x scale and two crops of the surrounding neighborhood — stacked into a 224x224x21 tensor. An ImageNet-pretrained ResNet-50, fine-tuned as a cell-type classifier and ensembled over five random initializations, converts that tensor into a morphology representation; its cell-type probability vector is concatenated with the seven mean expression values and passed to an XGBoost regressor that jointly predicts the remaining biomarkers. Training used 308 samples from 81 patients with head and neck squamous cell carcinoma at the University of Pittsburgh Medical Center, evaluated on held-out coverslips. Imputation reaches a mean patchwise Pearson correlation of 0.534 and a weighted cell-typing F1 of 0.727, versus 0.474 and 0.667 for an expression-only model; immune markers such as CD4, CD20, and CD45 exceed 0.70 correlation. Transferred without retraining, the same model scores 0.489 PCC and 0.614 F1 on a Stanford colorectal cohort where a locally trained reference model scores 0.583 and 0.605, and 0.475 PCC and 0.757 F1 on a Stanford head and neck cohort. Training on a single coverslip — 24% of the data — preserves cell-typing performance.

#Applications

7-UP suits translational oncology groups and pathology cores that already run low-plex clinical immunofluorescence and want CODEX-grade phenotyping without CODEX-grade cost. It enriches archived 7-plex material, supports tumor-microenvironment studies at cohort scale, and can extend a CODEX experiment itself past 40 measured markers by imputing on top of what was assayed.

#Impact

7-UP established that morphology extracted by a deep network is a usable substitute for unmeasured protein channels in multiplexed tissue imaging, and that the substitution survives all the way to clinical endpoints rather than stopping at correlation plots. It opened Enable Medicine's line of in-silico staining models, whose later entries move the same idea onto plain H&E. The limits are concrete: input and output are locked to a specific CODEX panel, per-marker accuracy varies widely and degrades further at a third clinical site (0.394 PCC on the Dana-Farber cohort), and the released GitLab repository contains training code but no checkpoint, license, or inference entry point, so the trained model itself has never been distributed.