Abstract A010: Benchmarking 3D against 2D deep learning-driven image-based profiling for predicting biological relationships amongst genes

Abstract Understanding how genes and proteins interact within the cell is essential for identifying drug targets and decoding signalling pathways. Recent advances, including foundation models such as OpenPhenom, Phenom-1 (Kraus et al., 2024) and Phenom-2 (Kenyon-Dean et al., 2024), have demonstrated the power of deep learning on 2D cell images to predict these biological relationships. However, such approaches remain limited by their inability to capture the full complexity of three-dimensional cell structure. Here, we benchmark a 3D morphology-based deep learning pipeline—Sentinal4DOne—against these state-of-the-art 2D imaging approaches, demonstrating a significant performance advantage in recovering known biochemical networks.Sentinal4DOne applies MorphoMIL, a multiple-instance learning framework recently published in Cell Systems (De Vries et al., 2025; PMID: 40112779), to a new 3D imaging dataset of over 35,000 melanoma cells subjected to RNA interference targeting genes from the Rho GTPase signalling axis, including RhoGEFs, RhoGAPs, and Rho-family GTPases. Using oblique plane light-sheet microscopy, single-cell shapes were quantified as point clouds and processed through a Dynamic FoldingNet encoder to extract geometric features. These were passed to the MorphoMIL model to generate compound-specific morphological embeddings. For each gene knockdown, we measured similarity to small-molecule phenotypes and calculated a profile which was the similarity to 8 different chemical compounds. These profiles were then used to predict interaction networks by ranking high-similarity gene pairs.To evaluate performance, we calculated recall scores against curated gold-standard databases including CORUM, Reactome, SIGNOR, and STRING, following the benchmarking strategy in Schubert et al., PLOS Comput Biol, 2023. Sentinal4DOne achieved consistently higher recall than all competing approaches. Compared to CellProfiler-derived features from the JUMP Cell Painting cp0016 dataset, our 3D morphological features achieved over 2× improvement in recall across all databases. Moreover, Sentinal4DOne outperformed Recursion’s OpenPhenom model, PerkinElmer’s Columbus platform, and our earlier 2D-based model, SentinalZero (Bousgouni et al., 2022; PMID: 36039362).These results highlight the power of 3D morphology as a biologically meaningful descriptor of perturbation. Not only do 3D shape-based models better capture heterogeneous cell states, they also enable more accurate inference of functional and physical gene relationships. By mapping protein networks from imaging data with higher precision, Sentinal4DOne represents a significant advance in phenotypic screening and target discovery pipelines. Citation Format: Matt De Vries, Reed Naidoo, Vicky Bousgouni, Georgia Mitsi, Chris Bakal. Benchmarking 3D against 2D deep learning-driven image-based profiling for predicting biological relationships amongst genes [abstract]. In: Proceedings of the AACR Special Conference in Cancer Research: Artificial Intelligence and Machine Learning; 2025 Jul 10-12; Montreal, QC, Canada. Philadelphia (PA): AACR; Clin Cancer Res 2025;31(13_Suppl):Abstract nr A010.

Récupéré en direct depuis OpenAlex et désinversé. Les résumés ne sont pas conservés dans cette base de données : les index inversés représentent 8,6 Go des 9,3 Go de texte de la base, et le serveur dispose de 13 Go libres.