Background: Novel cancer therapeutic antibodies targeting tumor-specific antigens are in development. However, precise quantification of target protein expression on Tumor Cells (TC) from immunohistochemistry (IHC) is challenging. Here, we developed an artificial intelligence powered image analyzer, called Universal IHC (UIHC), which robustly detects and quantifies targets of interest expressed in multiple untrained cancer types and antibodies.

Methods: UIHC model can detect TCs and classify IHC positivity across many antibody-cancer pairs, called “domains”. UIHC is trained on four domains: lung cancer (569K/107K) (No. of negative TC/positive TC), bladder cancer (270K/116K), and breast cancer stained with PD-L1 22C3 pharmDx (270K/96K), and breast cancer stained with HER2 4B5 (307K/256K); all annotated by board-certified pathologists. For comparison, we developed four AI models using only a single domain (SD). All models were evaluated on hold-out test sets divided in three groups: “in-domain” included the four training domains; “near-out-domain” contained PD-L1 SP142 and SP263 on lung cancer; “far-out-domain” contained 7 antibodies on 15 cancer types.

Results: For cell detection, the harmonic means of the precision and recall (F1 score) averaged on TC- and TC+ (mF1), formatted as (UIHC; avg-SD [min, max]) were: “in-domain” (72.5%; 68.9% [67.1%-69.9%]), “near-out-domain” (64.8%; 61.6% [59.7%-65.2%]), and “far-out-domain” (70.3%; 62.5% [59.0%-65.9%]).

At image level, cutoffs at 1% and 50% of TPS are applied for PD-L1 22C3 and SP263, thus 3-way accuracy is computed*. In far-out-domain, the mean TPS per antibody is used as a cutoff.

UIHC outperforms SD on most antibodies (Table 1).

Conclusion: We demonstrated that a UIHC, built by merging multiple domains during development, can be effective on untrained domains. UIHC can be applied to novel targets in future clinical research with minimal or no adaptation needed.

Model performance of the proposed UIHC model and four SD models. *3-way acc 1%-50% TPS cutoff

Hold-out test-setCell detection performance
(mF1)
Tumor Proportion Score (TPS) classification performance
(accuracy)
Test set group Staining(antibodies) Cancer type Number of negative TC/
positive TC 
UIHC model SD models
Average
[min-max] 
UIHC model SD models
Average
[min-max] 
In-domain test set PD-L1 22C3 Lung 37716/
13792 
76.7% 71.5%*
[65.6%-75.9%] 
80.8%* 77.3%*
[75.7%-78.7%] 
Near-out-domaintest set PD-L1 SP263 Lung 524/
485 
64.1% 64.0%
[57.6%-72.6%] 
80.0%* 72.5%*
[60.0%-80.0%] 
Far-out-domain test set TROP2 Pan-cancer 1243/
3289 
78.9% 70.7%
[63.7%-78.4%] 
94.0% 87.8%
[87%-90%] 
 MET  3493/
1134 
80.5% 77.7%
[75.1%-80.4%] 
96.0% 96.0%
[92.0%-100.0%] 
 Claudin18.2  2927/
74 
72.7% 63.4%
[55.1%-73.3%] 
94.0% 93.8%
[81.3%-100.0%] 
 DLL3  2973/
68.9% 42.5%
[40.3%-43.8%] 
100% 92.4%
[87.5%-94.0%] 
 HER3  2768/
138 
60.9% 54.2%
[44.0%-71.3%] 
94.0% 93.9%
[87.5%-88.0%] 
 FGFR2  2358/
438 
50.8% 58.9%
[52.8%-58.9%] 
81.0% 87.9%
[87.5%-88.0%] 
 Ecadherin  443/
912 
79.2% 74.6%
[72.1%-76.7%] 
90.0% 82.5%
[80.0%-90.0%] 
Hold-out test-setCell detection performance
(mF1)
Tumor Proportion Score (TPS) classification performance
(accuracy)
Test set group Staining(antibodies) Cancer type Number of negative TC/
positive TC 
UIHC model SD models
Average
[min-max] 
UIHC model SD models
Average
[min-max] 
In-domain test set PD-L1 22C3 Lung 37716/
13792 
76.7% 71.5%*
[65.6%-75.9%] 
80.8%* 77.3%*
[75.7%-78.7%] 
Near-out-domaintest set PD-L1 SP263 Lung 524/
485 
64.1% 64.0%
[57.6%-72.6%] 
80.0%* 72.5%*
[60.0%-80.0%] 
Far-out-domain test set TROP2 Pan-cancer 1243/
3289 
78.9% 70.7%
[63.7%-78.4%] 
94.0% 87.8%
[87%-90%] 
 MET  3493/
1134 
80.5% 77.7%
[75.1%-80.4%] 
96.0% 96.0%
[92.0%-100.0%] 
 Claudin18.2  2927/
74 
72.7% 63.4%
[55.1%-73.3%] 
94.0% 93.8%
[81.3%-100.0%] 
 DLL3  2973/
68.9% 42.5%
[40.3%-43.8%] 
100% 92.4%
[87.5%-94.0%] 
 HER3  2768/
138 
60.9% 54.2%
[44.0%-71.3%] 
94.0% 93.9%
[87.5%-88.0%] 
 FGFR2  2358/
438 
50.8% 58.9%
[52.8%-58.9%] 
81.0% 87.9%
[87.5%-88.0%] 
 Ecadherin  443/
912 
79.2% 74.6%
[72.1%-76.7%] 
90.0% 82.5%
[80.0%-90.0%] 

Citation Format: Biagio Brattoli, Mohammad Mostafavi, Sangjoon Choi, Taebum Lee, Seokhwi Kim, Wonkyung Jung, Soo Ick Cho, Jinhee Lee, Keunhyung Chung, Jeongun Ryu, Seonwook Park, Sergio Pereira, Seunghwan Shin, Chan-Young Ock. Universal immunohistochemistry positivity classification of cancer cells across multiple cancer types and antibodies using artificial intelligence. [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2023; Part 1 (Regular and Invited Abstracts); 2023 Apr 14-19; Orlando, FL. Philadelphia (PA): AACR; Cancer Res 2023;83(7_Suppl):Abstract nr 5392.