CellForge Framework: Automated Cell Modeling

• CellForge Framework is a comprehensive system designed for automated cell modeling, simulation, and quantitative analysis of cellular processes.
• Its architecture features modular Task Analysis, Method Design, and Experiment Execution modules that leverage multi-agent consensus and JSON-RPC communication.
• The framework enables accurate virtual cell modeling, morphometric cell analysis, and efficient deep learning integration to advance computational cell biology.

The CellForge Framework is a multi-component system for automated modeling, simulation, and quantitative analysis of biological cells and cellular processes. It integrates agentic model construction, spatial hybrid systems simulation, and advanced feature extraction for cellular image data, thereby enabling comprehensive analyses ranging from virtual perturbation predictions to morphometric cell characterization.

1. Architecture and System Design

CellForge is constructed as an autonomous end-to-end workflow that receives raw single-cell multi-omics data and research task specifications, and outputs optimized virtual cell models alongside executable code for model training and inference (Tang et al., 4 Aug 2025). The framework is composed of three primary modules:

• Task Analysis: Performs dataset profiling and literature retrieval, leveraging a multi-layered search strategy (alternating breadth-first and depth-first traversal) and relevance scoring via cosine similarity of Sentence-BERT embeddings.
• Method Design: Implements a multi-agent, graph-based discussion among specialized agents (Data Analyst, Problem Investigator, Baseline Assessor, Deep Learning Expert, etc.), moderated by a Critic Agent. A consensus protocol using confidence scores

$$c_t^{(i)} = \lambda_1 c_{t-1}^{(i)} + \lambda_2 \text{CriticAgentScore}(m_t^{(i)}) + \lambda_3\frac{1}{k-1} \sum_{j\neq i} \text{PeerScore}(m_t^{(i)}, E^{(j)})$$

iteratively refines candidate modeling strategies until convergence.

• Experiment Execution: Automatically generates and self-debugges code, supports model training with early stopping, adaptive learning rate schedules, and checkpointing.

Communication between modules uses JSON-RPC with persistent shared memory, enabling complex reasoning and traceability.

2. Modeling and Simulation Methodology

In its simulation core, CellForge leverages a domain-specific modeling language (Mechanica Modeling Language, MML) tailored to biological domain concepts (Somogyi et al., 2017). The system features:

• Objects and Processes Abstraction: Objects (particles, material regions, chemical concentrations) encode stateful entities, while processes describe transformations, including chemical reactions, diffusions, and mechanical force transfer.
• Spatial Data Representations: Native primitives (particles, links, spatial fields, material regions) define intra- and extra-cellular geometries, with particles holding spatial position, mass, radius, and other physical attributes. Links dynamically encode mechanical bonds that can form or break.
• Coupled Interactions: Unified representation of chemical, mechanical, and electrical phenomena. Chemical networks are instantiated as processes with notation akin to conventional chemical equations; mechanical couplings are defined via dynamic force terms.
• Simulation Algorithms: Compiler parses MML via recursive descent and semantic analysis, generating two coupled dynamical systems: ODEs for reaction network dynamics and Lagrangian equations for particle dynamics. ODEs follow

$$\frac{dC}{dt} = [ N \cdot \nu(C, r, v, p);\: f(C, r, v, p) ]$$

where $N$ is a stoichiometry matrix, $\nu$ is the transformation rate function, $f$ denotes additional rate processes.

Mechanical evolution is integrated mesh-free, using force balance:

$$\frac{d^2 r_i}{dt^2} = \frac{1}{m_i} \left[\sum_{j \neq i} \left(F^L_{ij} + F^C_{ij} + F^D_{ij} + F^R_{ij}\right) + F^{ext}_i\right]$$

Continuous variables are stored in contiguous memory compatible with CVODE; interparticle force computations employ mdcore, utilizing Verlet and cell lists.

3. Feature Extraction and Morphometric Analysis

CellForge incorporates advanced image analysis via integration with feature extraction methodologies like Cellpose+ (Huaman et al., 24 Oct 2024). Following deep learning-based segmentation, quantitative morphological metrics are derived:

Metric	Formula	Biological Interpretation
Roundness ($R$)	$R = \frac{4\pi a}{p^2}$, $a$ = area, $p$ = perimeter	Sphericity/deviation from circularity
Cytoplasm/Nucleus Ratio ($C_i/N_i$)	Computed per cell	Differentiation or viability marker
Area Coverage	$\frac{\sum A_i}{A_t}$, $A_i$ = area, $A_t$ = image area	Density and confluence
Voronoi Entropy ($S_\text{Vor}$)	$- \sum p_i \log p_i$, $p_i$ = polygon proportion	Spatial cell organization
Continuous Symmetry Measure (CSM)	$\frac{\sum |M_i - \hat{M}_i|^2}{n S_i}$	Departure from ideal symmetry

Metrics are computed after segmentation and utilize libraries such as DIP and scipy. Masks are stored in .npy files for post-hoc manual refinement.

4. Agentic Consensus and Automated Model Generation

The core innovation of CellForge lies in its agentic model design system (Tang et al., 4 Aug 2025). Multiple agents, each representing a unique role (Data Engineer, Architecture Expert, Deep Learning Practitioner, etc.), engage in a collaborative, graph-structured dialogue. Each agent:

• Proposes model architectures using its specialized perspective
• Critiques and scores proposals from other agents and the central Critic Agent
• Updates its confidence score per round using the aforementioned consensus formula

Discussion rounds iterate until all experts’ confidence ratings pass a specified consensus threshold, minimizing variance across agent scores. This collaborative process demonstrably yields model architectures superior to those generated by single-agent or one-shot approaches.

5. Biological and Computational Applications

CellForge supports diverse biological modeling and analysis applications:

• Virtual Cell Modeling: Predictive modeling across gene knockouts, drug treatments, and cytokine stimulations in scRNA-seq, scATAC-seq, and CITE-seq datasets. Benchmark results include MSE of 0.0051 (±0.0063), PCC of 0.9883 (±0.0459), and $R^2$ of 0.9761 (±0.0803) in gene knockout tasks; up to a 20% PCC improvement over ChemCPA for drug perturbations.
• Morphometric Cell Analysis: Automated quantification of cellular morphology and organization for biocompatibility assessment, differentiation potential, and viability screening. Data from DAPI (nuclear) and FITC (cytoplasmic) stained fibroblast images were used to calibrate segmentation and feature extraction.
• Mechanical and Chemical Coupling Simulations: Chemotaxis models simulating cellular responses to gradients and dynamic adhesion (Velcro-like mechanisms), reaction network simulations, and spatial connectivity in tissue modeling.

6. Performance, Accuracy, and Code Accessibility

CellForge’s output models and code consistently exceed baseline and state-of-the-art approaches in predictive and morphometric analysis accuracy. Manual refinement of segmentation outputs minimizes deviation from expert counts (Huaman et al., 24 Oct 2024). Qualitative and quantitative expert and LLM "judge" evaluations confirm superior plan generation and implementation (Tang et al., 4 Aug 2025).

Production code is publicly available:

• https://github.com/gersteinlab/CellForge

Utilization requires proficiency in PyTorch, installation of Python packages per requirements.txt, and compatibility with AnnData/h5ad formats. Detailed instructions and usage examples are provided in the repository README.

7. Future Extensions and Ecosystem Integration

Planned enhancements for CellForge and its associated modules include:

• Expansion of extractable features (e.g., new shape descriptors, transition to 3D segmentation) and broader analytical scope (Huaman et al., 24 Oct 2024).
• Modular GUI interfaces for customized plug-in analyses.
• Support for diverse staining and imaging modalities.
• Further integration of deep learning models to increase robustness to imaging artifacts.

These extensions are aimed at yielding a more modular, scalable pipeline integrating cell segmentation, morphometric profiling, downstream statistical analyses, and virtual cell model inference in a unified system.

---

CellForge defines an integrated methodology for iterative agentic model design, biophysically motivated simulation, and high-throughput cell image analysis, providing a comprehensive and extensible platform for research in computational cell biology (Tang et al., 4 Aug 2025, Somogyi et al., 2017, Huaman et al., 24 Oct 2024).