Abstract Multiplex immunofluorescence (mIF) is a powerful tool for profiling dozens of biomarkers from a single tissue section. Customized mIF panels enable oncology researchers to phenotype cells within the tumor microenvironment, interrogate activation status of immune cells, quantify expression levels of biomarker targets, and explore the spatial organization of cells. Yet, analysis of large mIF datasets remains a bottleneck, largely due to the difficulty of accurately phenotyping single cells across large tissues. The incorporation of deep-learning algorithms into mIF analysis pipelines has helped overcome some limitations of traditional intensity gating by using stain morphology to add robustness to intensity variation, spatial spillover, and tissue artifacts. However, these deep-learning algorithms are costly to develop from a data volume and/or annotation perspective, often requiring fine-tuning on manual labels from target datasets to achieve acceptable performance. Therefore, there is demand for algorithms that efficiently generalize across batches, mIF panels, and tissue types. Here, we present a label-free framework for adapting a pretrained single-channel feature extractor into a mIF whole panel classifier capable of zero-shot cell phenotyping. Our source model is a single-channel feature extractor trained on 20 million annotations spanning over 40 biomarker classes, with text conditioning to encode marker-specific interpretations. We adapt this model into a whole panel classifier by introducing both single-marker binary heads and a multi-marker phenotyping head, jointly trained through self-distillation to enforce per-channel consistency while promoting information maximization across the multiplexed panel. This strategy preserves interpretable single-marker classification while leveraging cross-channel context for accurate and scalable phenotyping of mIF datasets. We applied this framework to 27 FFPE samples stained with a 17-marker mIF TME panel using the PaletrraTM (NeoGenomics Laboratories, Inc) platform. The samples are from a cohort of metastatic melanoma patients treated with pembrolizumab, with our analysis revealing significant differences in immune populations between non-responders and responders. Notably, we showed that our framework achieved higher cell phenotyping accuracy and reduced errors from common mIF issues such as batch effects and spatial spillover, as compared to other mIF image analysis techniques. Overall, our method provides a lightweight, automated method for adapting pretrained mIF algorithms to new panels with strong zero-shot performance to novel biomarkers. Citation Format: Kevin Gallagher, Jiong Fei, Judy Kuo, Maryam Rohafza, Mitchell P. Levesque, Julia M. Martínez-Gómez, Marianne Thio, Erinn A. Parnell, Qingyan Au, Harry Nunns. PaletrraTM AI: Automated phenotyping of multiplex immunofluorescence datasets via information maximizing self-training abstract. In: Proceedings of the American Association for Cancer Research Annual Meeting 2026; Part 1 (Regular Abstracts); 2026 Apr 17-22; San Diego, CA. Philadelphia (PA): AACR; Cancer Res 2026;86(7 Suppl):Abstract nr 1456.
Gallagher et al. (Fri,) studied this question.