AI-augmentation of multiplex chromogenic assays for use with colocalized multi-specific targets.

e15178 Background: Assessing the co-localization of protein targets is important to aid in the selection of patients who may benefit from emerging multispecific T-cell engagers and antibody drug conjugates; however the current gold standard of pathologist interpretation of individual IHC stains is suboptimal. Separate IHC stains on consecutive sections do not allow evaluation of staining on the same tumor cells, making assessment of true co-localization impossible. Multiplexed chromogenic IHC (mIHC) immunostains have been developed to overcome these challenges, though these have proven difficult for pathologists to interpret, especially when one of the antigens is highly expressed and the other antigen is expressed at low levels. Machine learning techniques are now available to augment pathologist interpretation of immunostains, and we have developed a suite of tools specifically to augment the interpretation of mIHC for protein colocalization. Methods: To develop a novel algorithm to augment single-slide mIHC assays measuring colocalized tumor targets we developed a suite of deep learning algorithms to enable convolutional neural net deep learning approaches. These tools included color enhancement and virtual DAB staining built in reference to sandwiched IHC staining on serial sections with traditional monoplex IHC stains. These tools allow for direct annotation and deep learning of individual antibody stains within a mIHC slide, and development of models for individual and double positivity at cellular resolution. Combining these models with existing cell detection and classification models enables direct granular assessment of the performance of mIHC to colocalize targets within on-target cell types. Results: Development of the AI-augmented mIHC showed high performance characteristics compared to sandwich IHC for detecting individual targets presence on tumor cells ( > 80% balanced accuracy & F1). Additionally, the development of virtual single-plex DAB and color enhancement tools showed improvement in Pathologist precision and recall when compared to scoring presence of a target on the original mIHC image. Finally, precision in calculating co-expression was observed to improve with AI-augmentation compared to pathologist assessment without access to mIHC AI tools. Conclusions: As the assessment of co-localization of protein targets becomes essential, pathologists will need tools to aid them in the reproducible evaluation of mIHC slides. We have demonstrated that deep learning algorithms provide a reliable and reproducible methodology to aid pathologists.