A generative AI framework for disease-specific lung microtissue bioengineering

1. Problem Statement

Chronic lung diseases such as Chronic Obstructive Pulmonary Disease (COPD) and Pulmonary Fibrosis (PF) pose significant global health burdens. Current research and therapeutic development face critical limitations:

• Model Limitations: Traditional in vitro models lack complex 3D tissue architecture, while animal models often fail to translate to human physiology due to biological differences and ethical concerns.
• Organ Shortage: Lung transplantation is the only curative option for end-stage disease but is severely limited by donor shortages and graft rejection.
• Biofabrication Gaps: While 3D bioprinting can create scaffolds, existing methods often rely on simplified geometries or lack the ability to recapitulate the specific, complex micro-architectures of diseased lung tissue (e.g., fibrotic scarring or emphysematous alveolar destruction).
• Integration Gap: There is a lack of a unified pipeline that connects high-resolution biological imaging, AI-driven generative design, and high-resolution biofabrication to create de novo disease-specific tissue models.

2. Methodology: The GLAM Framework

The authors present Generative Lung Architecture Modeling (GLAM), an integrated pipeline that closes the loop between biological imaging, AI design, and tissue engineering. The workflow consists of four main stages:

A. Biological Data Acquisition (In Vivo to Ex Vivo)

• Disease Models: Murine models were used to generate three conditions: Healthy (control), Pulmonary Fibrosis (induced by bleomycin), and Emphysema/COPD (induced by elastase or cigarette smoke).
• Precision-Cut Lung Slices (PCLS): Lungs were explanted, embedded in agarose, and sectioned into 300 µm thick slices to preserve native 3D architecture and extracellular matrix (ECM).
• Processing: PCLS were decellularized to remove cellular components (reducing light scattering) and stained for ECM proteins (Collagen I and Sirius Red).
• Imaging: High-resolution 3D confocal microscopy was used to acquire 60 volumetric datasets (z-stacks) representing the native ECM topology of healthy, fibrotic, and emphysematous tissues.

B. Computational Pipeline & AI Generation

• Preprocessing: Raw image stacks were tiled into 200 µm cubes, interpolated to isotropic resolution, and segmented using a pre-trained PlantSeg U-Net model to generate probability maps of the ECM.
• Generative Modeling: A custom 3D Diffusion Model (based on a U-Net architecture with 3D spatial attention layers) was trained on the segmented probability maps.
  • Three separate models were trained for each condition (Healthy, Fibrotic, Emphysematous).
  • The models learned disease-specific structural patterns to synthesize novel, realistic 3D digital meshes.
• Mesh Generation: Both original segmented data and AI-generated synthetic volumes were converted into printable 3D meshes using the Marching Cubes algorithm, followed by post-processing (hole filling, normal fixing) to ensure watertight geometries.

C. High-Resolution Biofabrication

• Printing Technology: Two-photon stereolithography (2PP) was used to print the digital meshes.
• Bioink: Gelatin-methacryloyl (GelMA/GM10) hydrogel was used as the bioink, providing biochemical cues for cell adhesion.
• Parameters: Printing was performed with a 500 nm hatching distance to ensure structural integrity of fine alveolar features.

D. Biological Validation

• Colonization: The printed scaffolds were seeded with human MRC-5-GFP fibroblasts.
• Analysis: Cell attachment, spreading, and distribution were assessed via 2-photon fluorescence microscopy after 24 hours.

3. Key Contributions

• First End-to-End GLAM Pipeline: Demonstrates a fully integrated workflow from in vivo disease modeling to de novo AI-designed tissue fabrication.
• Disease-Specific Generative AI: Successfully trained diffusion models to synthesize anatomically accurate 3D lung architectures for distinct pathological states (fibrosis and emphysema), not just healthy tissue.
• Validation of Synthetic Data: Provided a rigorous multi-tiered validation framework (FID, Precision/Recall, MS-SSIM, Euclidean distance transforms) proving that AI-generated meshes statistically match native tissue distributions.
• Biocompatible Microtissues: Successfully fabricated and colonized 200–300 µm lung microtissues, demonstrating that AI-generated geometries can support human cell attachment and migration.

4. Key Results

• Model Performance:
  • The diffusion models achieved low validation losses and generated foreground voxel ratios closely matching the training baselines (e.g., Healthy: 0.258 vs. 0.247 baseline).
  • FID Scores: 35.5 (Emphysema), 45.6 (Healthy), and 47.9 (Fibrotic), indicating high fidelity to the statistical distribution of real microscopy data.
  • Diversity: Low Multi-Scale Structural Similarity (MS-SSIM) values (0.05–0.07) confirmed the models generate diverse, non-redundant structures.
• Fabrication Success:
  • Both tissue-derived and AI-generated meshes were successfully printed with high shape fidelity (340x340x140 µm and 200x200x200 µm blocks, respectively).
  • Printed scaffolds accurately replicated alveolar septal boundaries and ECM wall structures.
• Biological Compatibility:
  • Fibroblasts successfully attached and spread along the printed scaffolds.
  • Observation: Model-generated scaffolds showed slightly reduced cell density compared to tissue-derived ones, likely due to smaller scaffold volume and minor topological artifacts (small enclosed alveolar regions <140 µm³) that limited cell access.
  • Emphysematous scaffolds showed the lowest cell attachment, correlating with enlarged alveolar diameters and reduced surface complexity.

5. Significance and Future Outlook

• Regenerative Medicine: This framework offers a pathway to create patient-specific or disease-specific tissue surrogates, potentially addressing the organ donor shortage by enabling the fabrication of functional lung tissue.
• Drug Discovery & Preclinical Testing: The ability to generate scalable, customizable, and disease-specific microtissues provides a superior platform for drug screening and studying disease mechanisms compared to traditional 2D cultures or animal models.
• Reduced Animal Use: By using AI to generate synthetic training data and tissue models, the need for extensive animal experimentation can be reduced.
• Limitations & Future Work:
  • Scale: Current 2PP printing is limited to small volumes (<1 mm³); future work requires scaling up via volumetric printing.
  • Complexity: Current models focus on ECM; future iterations must integrate vascular networks and multiple cell types (epithelial, endothelial) for functional gas exchange.
  • Safety: Rigorous validation is required to prevent "hallucinations" (biologically implausible structures) in AI-generated medical designs before clinical translation.

In conclusion, the GLAM framework establishes generative AI as a functional design layer for tissue engineering, bridging the gap between digital imaging and physical biofabrication to create realistic, disease-specific lung microtissues.