A Digital Twin-Inspired Closed-Loop Latent Simulation Framework for Cross-Cohort Breast Cancer Subtype Classification...

Area of Science:

• Computational Biology and Bioinformatics
• Machine Learning in Oncology
• Medical Image Analysis

Background:

• Current deep learning models for breast cancer PAM50 subtype classification use a single-pass prediction, limiting iterative refinement and uncertainty analysis.
• Integrating diverse data modalities like histopathology, transcriptomics, and mammography presents challenges due to missing data across cohorts.
• Existing methods often rely on per-patient data fusion, which is not feasible when patients lack complete multimodal data.

Purpose of the Study:

• To develop an advanced deep learning framework, the Cross-Cohort Modality-Disjoint Latent Simulation (CDLS), for improved breast cancer subtype classification.
• To enable iterative representation refinement and uncertainty trajectory analysis within a closed-loop system.
• To effectively integrate heterogeneous data from multiple cohorts under a modality-disjoint regime.

Main Methods:

• Implemented a Cross-Cohort Modality-Disjoint Latent Simulation (CDLS) framework integrating histopathology (WSI), transcriptomics (RNA-seq), mammography, and clinical data.
• Employed a Proximal Policy Optimization (PPO)-governed stochastic policy to refine a 7-dimensional latent state ($z$) over 5 optimization steps using a Twin-GRU transition model.
• Incorporated a closed-loop latent feedback mechanism involving k-Nearest Neighbors (kNN) retrieval for aligning simulated states with real patient embeddings.

Main Results:

• Achieved a Balanced Accuracy of 0.870 ± 0.044 and MCC of 0.904 ± 0.046 in multi-seed evaluations (n=4).
• Cross-validation confirmed stability with Accuracy 0.871 ± 0.029 and MCC 0.897 ± 0.038.
• The chosen 7-dimensional latent space bottleneck was guided by intrinsic dimensionality estimates (d_id = 6.3 ± 0.4).

Conclusions:

• The CDLS framework offers a robust approach for breast cancer subtype classification by enabling iterative representation refinement and handling modality-disjoint data.
• The closed-loop latent feedback mechanism effectively aligns simulated and real patient representations, enhancing classification performance.
• PPO selection was based on trajectory geometry properties, ensuring broader latent coverage and path diversity, crucial for complex biological systems.