PhysiBench: Multiscale Biology Benchmark
- PhysiBench is an open multiscale systems biology benchmark featuring 612 executable intracellular Boolean network variants and 120,000 multiscale simulation runs.
- The benchmark standardizes heterogeneous models via a three-stage pipeline (mutation, online filtering, and offline sensitivity analysis) to ensure controlled, reproducible simulation outputs.
- It supports surrogate modeling, data-driven inference, and simulation-based optimization, offering a robust platform for comparing computational methods in systems biology.
“PhysiCa-Bench” is not the official name used in the primary publication most closely associated with the term. In that work, the resource is consistently named PhysiBench, and “PhysiCa-Bench” is best understood as a typo or informal alias referring to the same benchmark. PhysiBench is an open, controlled benchmark resource for systems and computational biology composed of 612 executable intracellular Boolean regulatory network variants and a paired dataset of 120,000 time-resolved multiscale stochastic simulations generated in the PhysiBoSS/PhysiCell framework (Masera et al., 16 Jun 2026). Its design emphasizes executable reproducibility, controlled structural and behavioral diversity, and standardized multiscale coupling between intracellular Boolean dynamics and agent-based cellular simulations.
1. Nomenclature and disambiguation
Within the systems-biology literature, the relevant resource is PhysiBench, not “PhysiCa-Bench.” The paper explicitly presents PhysiBench as an open benchmark suite and dataset for developing and evaluating computational methods in systems biology, and there is no occurrence of “PhysiCa-Bench” as an official name (Masera et al., 16 Jun 2026). The benchmark is centered on executable intracellular Boolean regulatory networks coupled to multiscale simulations, rather than on generic physical simulation or large-language-model evaluation.
This naming point matters because similarly suggestive labels appear in unrelated areas. A separate 2025 benchmark in fundamental physics introduces a “Large Physics Benchmark” or “Physics Creativity and Reasoning Benchmark” for evaluating LLMs on multiple-choice, derivational, and code-based physics tasks; that paper does not use “PhysiCa-Bench” as an established alternate name either (Barman et al., 29 Jul 2025). The two resources differ in domain, task structure, and evaluation target: PhysiBench addresses multiscale systems biology simulation, whereas the physics benchmark addresses LLM reasoning and creativity in fundamental physics.
A common misconception is therefore to treat “PhysiCa-Bench” as a standardized, official title spanning several domains. The available papers do not support that interpretation. In the biological setting, the term maps to PhysiBench; in the physics-LLM setting, it does not denote the official project name.
2. Benchmark composition and biological scope
PhysiBench consists of two tightly linked components: a benchmark suite of 612 executable intracellular Boolean regulatory network variants and a simulation dataset of 120,000 independent multiscale runs, organized as 60 selected models × 2,000 contexts/model (Masera et al., 16 Jun 2026). The suite was created to support direct simulation, surrogate modeling, data-driven inference, simulation-based optimization, and comparative benchmarking under well-defined inputs and outputs.
The Boolean models originate from seven published PhysiBoSS/MaBoSS-compatible networks spanning several biological domains: mammalian cell-cycle restriction-point control, the Drosophila segment-polarity network, EGF/TNF signaling, gastric cancer signaling, macrophage activation, prostate cancer signaling, and TNF-induced cell-fate decision. These source networks collectively cover cell-cycle control, developmental patterning, cancer signaling, immune response states, and cell-fate decisions (Masera et al., 16 Jun 2026).
Each source model is standardized to a common interface with one input node for external stimulation and three output nodes coupling intracellular state to survival, proliferation, and death behaviors in the multiscale simulator. The models are also validated syntactically for MaBoSS (.bnd/.cfg) and PhysiBoSS execution. This standardization is central to the benchmark’s comparability: different biological families are exposed through the same simulator-level interface, which makes both on-demand simulation and cross-model method evaluation more controlled.
| Component | Quantity | Description |
|---|---|---|
| Benchmark variants | 612 | Executable intracellular Boolean regulatory network variants |
| Selected models for trajectories | 60 | Models used for large-scale simulation dataset generation |
| Multiscale simulations | 120,000 | Independent time-resolved stochastic runs |
A plausible implication is that PhysiBench is structured not merely as a model repository but as a benchmark family with standardized executable semantics. This distinguishes it from resources focused primarily on model discovery or archival sharing.
3. Variant construction and filtering pipeline
The 612 retained variants are generated through a three-stage pipeline: mutation-based model construction, online behavioral filtering, and offline sensitivity filtering (Masera et al., 16 Jun 2026). Mutation is stochastic and local, with input and output nodes protected. The operators used are switch_nodes_logic, replace_logical_operator, replace_node_inside_logic, negate_subexpression, add_input_to_logic, add_new_node, randomize_node_logic, and randomize_parameter, with randomize_parameter disabled in the released generation by setting its probability to 0.
The mutation process is bounded by explicit safeguards. The maximum number of added nodes per lineage is 45, the maximum network size after interface standardization is 63 nodes, and mutation depth is adaptively scheduled between a minimum/initial value of 10 and a maximum of 2,000 operations. After acceptance, the mutation depth is multiplied by ; after rejection, it is multiplied by . The manuscript also provides pseudocode specifying that up to 200,000 candidates are tested.
Online behavioral filtering is designed to reduce redundancy. Each candidate model is simulated under 48 fixed protocols with a constant initial cell configuration. Its behavioral signature is formed by concatenating alive-cell counts from the final six recorded time points for each protocol. Similarity between candidate and accepted signatures is measured with Pearson correlation; a candidate is accepted only if its maximum correlation with the accepted pool is < 0.85, equivalently if “novelty” is > 0.15. This stage yields 2,122 candidates across families after online filtering (Masera et al., 16 Jun 2026).
Offline sensitivity filtering is then used to retain input-responsive models. Each candidate is simulated once across 215 contexts, with controlled variation in initial conditions and stimulation protocols. For each simulation, the pipeline computes nine output summaries centered on alive-cell count and spatial organization with respect to circular or square target regions, including shifted and distribution-weighted variants. Across contexts, it computes standard deviation (SD) and coefficient of variation (CV), and retains only models for which every summary satisfies SD > 30.0 and CV > 0.2. This stage produces the final 612-model benchmark suite.
4. Multiscale execution model and data organization
PhysiBench is executed in PhysiBoSS/PhysiCell, which integrates MaBoSS Boolean networks inside agent-based cells (Masera et al., 16 Jun 2026). Each cell contains an independent MaBoSS instance updated asynchronously in MaBoSS 2.0 style. The intracellular input node receives external stimulation from the PhysiCell microenvironment, parameterized by temporal schedule and boundary values, while the three standardized output nodes modulate survival, proliferation, and death in the agent-based layer.
The paper does not provide explicit update equations for Boolean dynamics such as , nor explicit microenvironment PDEs. It states that asynchronous update details follow MaBoSS conventions. It also does not provide formal sensitivity or graph-diversity formulas, although the graph distances used are specified.
The 120,000-run dataset is generated from 60 selected models under systematically sampled stimulation protocols and fixed model-level initial configurations. For each selected model, contexts are sampled independently from ten evenly spaced candidates for each parameter. The ranges are: stimulation duration 5–200 simulation-time units, stimulation period 5–800 simulation-time units, and each boundary value 0–10, subject to the constraint treatment_period ≥ treatment_duration. Valid protocols are shuffled, and 2,000 are retained per model (Masera et al., 16 Jun 2026).
The released files preserve a strict linkage between simulation identity, model identity, inputs, initial conditions, and outputs:
| Artifact | Contents |
|---|---|
variant_models_manifest.json |
reference_model, variant_ID, model_ID, bnd_file, cfg_file |
multiscale_simulations_manifest.json |
Simulation index linking parent model identifiers and simulation_ID |
initial_positions.json |
Fixed model-level initial configuration |
input_parameters_<simulation_ID>.json |
treatment_duration, treatment_period, TNF_dirichlet_xmin, TNF_dirichlet_xmax, TNF_dirichlet_ymin, TNF_dirichlet_ymax |
cell_data_<simulation_ID>.json.gz |
Time-resolved time, x_positions, y_positions, z_positions, current_phase |
Initial configurations are encoded with the fields type, center, density, cell_type, mode, and length, where type is circle or square, and length denotes radius or half-side length. Cell-level outputs record positions in micrometers and current_phase in the PhysiCell integer encoding: live = 14, apoptotic = 100, necrotic = 101/102/103 (Masera et al., 16 Jun 2026).
Reproducibility is specified operationally. A trajectory is reproduced by identifying the run through the quadruplet <reference_model, variant_ID, model_ID, simulation_ID>, loading model.bnd and model.cfg, applying the shared initial_positions.json, configuring the stimulus through input_parameters_<simulation_ID>.json, setting the recorded stochastic seed, executing PhysiBoSS/PhysiCell with fixed simulator settings for that model_ID, and verifying that the output matches cell_data_<simulation_ID>.json.gz at saved time points.
5. Validation, diversity analysis, and benchmark use
Technical validation covers both file integrity and executability. The released resource checks the 612 model directories for .bnd/.cfg presence, the 120,000 simulations for the presence of input_parameters and parsable compressed cell_data files, and both manifests for consistency and uniqueness. Representative PhysiBoSS runs were executed directly from the released files without per-model adaptation (Masera et al., 16 Jun 2026).
Structural diversity is assessed by converting Boolean rules to directed signed graphs and computing pairwise distances using DeltaCon, Ipsen–Mikhailov, and Quantum Jensen–Shannon divergence. Control tests show that shuffled node labels of a reference model yield low intra-family distances, whereas structurally different models yield higher distances. The reported within-to-between family ratios are 0.76–0.85, indicating substantial but traceable structural diversity.
Behavioral heterogeneity is quantified from multiscale outputs. The average pairwise correlation between output-response profiles across selected models is 0.06015, consistent with low redundancy. The outlier-strength distribution, using a modified z-score > 3.5 criterion during selection, spans orders of magnitude, with maximum approximately , mean 42.44, and quartiles 0.45, 6.40, 18.79 (Masera et al., 16 Jun 2026).
The benchmark is intended for several technical use cases. For surrogate modeling, the recommended task is to map context inputs to population-level outputs or spatial trajectories, with metrics such as RMSE/MAE on alive-cell counts, distributional distances for spatial organization, and temporal correlation on trajectories. Suggested data splits are 70/15/15 train/val/test over simulation_IDs within each model_ID, optionally stratified by treatment_period and duration to probe extrapolation. For data-driven inference, the nine output summaries are used for feature importance or sensitivity analysis via CV and SD. For simulation-based optimization, the controllable variables are treatment_period, treatment_duration, and the boundary values; objectives include minimizing alive-cell count or maximizing containment in target regions. For comparative benchmarking, the guidance is to fix seeds and simulator settings, evaluate on identical context subsets, report paired metrics, and use structural-diversity strata to probe robustness (Masera et al., 16 Jun 2026).
The dataset footprint is approximately 23 GB for the 120,000 compressed JSON trajectories, and <50 MB for the 612 model files. Large-scale runs are described as benefiting from HPC or containerized batch execution, and the project distributes scripts and a Snakemake workflow through the PhysiBench GitHub repository.
6. Limitations, comparisons, and prospective extensions
PhysiBench has explicit limitations. The variants are synthetic, generated by stochastic mutation and filtering rather than calibrated to experimental data, and are intended for method development, not biological prediction (Masera et al., 16 Jun 2026). The intracellular formalism is Boolean, so it captures discrete switch-like regulation rather than graded dose response or detailed kinetics. The spatial setting is 2D, without 3D tissue architecture or ECM mechanics, and out-of-plane diffusion is not represented. The stimulation protocol is an abstract input, not a mechanistic PK/PD model.
These constraints define the scope of valid interpretation. A common misuse would be to treat high performance on PhysiBench as evidence of biological calibration; the paper does not support that conclusion. A more defensible reading is that the resource provides controlled, executable multiscale testbeds for algorithmic comparison under fixed simulator semantics.
The benchmark is also situated among adjacent infrastructures. The paper explicitly compares it with PDEBench, WeatherBench, and The Well, which standardize simulation datasets and tasks in other domains, and with biology-oriented resources such as BioModels, BioSimulators/BioSimulations, PhysiBoSS-Models, and Boolean repositories, which emphasize model sharing and discovery rather than controlled benchmark families and paired multiscale trajectories (Masera et al., 16 Jun 2026). A related but distinct line of work is the extensible benchmark suite for learning to simulate physical systems introduced in 2021, which organizes spring, wave, spring mesh, and Navier–Stokes problems for objective evaluation of learning-based simulators rather than intracellular Boolean-agent-based biology (Otness et al., 2021). Likewise, the physics-community “Large Physics Benchmark” is an LLM evaluation framework, not a biological multiscale simulation benchmark (Barman et al., 29 Jul 2025).
Future directions proposed for PhysiBench include extending model families, adding graded or hybrid intracellular dynamics, moving to 3D simulations, enriching microenvironment models, increasing the number of inputs and outputs, and providing calibrated datasets and standardized downstream tasks and leaderboards (Masera et al., 16 Jun 2026). A plausible implication is that the benchmark is intended as infrastructure for methodological convergence in computational biology, analogous to how standardized benchmarks have functioned in other simulation and AI subfields, but within the specific constraints of executable PhysiBoSS/PhysiCell multiscale models.