Morpheus.AI: agents autonomously generate, execute and refine virtual tissue models
Listen
Radio episode about this paper
Transcript
Introduction to the show: ident: Genomics Radio. Generated commentary on the latest computational biology and genomics papers.
Ines: I'm Ines.
Marcus: And I'm Marcus, here with Yuki.
Ines: We're looking at the paper: "Morpheus.AI: agents autonomously generate, execute and refine virtual tissue models".
Marcus: Here's the gist of the paper. Virtual tissue models link observations to mechanistic hypotheses, but building them demands expertise in biology, mathematics and programming <ref:2611.05413#pg2>.
Ines: Here, large language model (LLM) agents paired with the simulator Morpheus (Morpheus.AI) autonomously built such models from fifteen benchmark studies, five without a published model <ref:2611.05413#pg3>.
Marcus: With a frontier LLM, agents that ran their simulations met 83–86% of study-specific criteria in under two hours and for under US money per model <ref:2611.05413#pg3>.
Ines: Such agents could make in silico modeling routine <ref:2611.05413#pg3>.
Paper summary: Ines: So to recap this paper "Morpheus.AI: agents autonomously generate, execute and refine virtual tissue models," the core thesis is that virtual tissue models connect biological observations to mechanistic hypotheses, but traditionally building them takes a lot of slow work requiring expertise in biology, mathematics and programming >
Marcus: The claim they make is that using large language model agents paired with their simulator Morpheus.AI allows these agents to autonomously build those virtual tissue models from fifteen benchmark studies, including five studies where there was no published model to start with >
Yuki: What this means for us is that it tackles the problem of building those multiscale virtual tissue models, which used to be described as slow and specialist work needing deep abstraction into assumptions and casting them in a mathematical framework >
Ines: And they show that these agents can handle the complexity, because they call tools, run their own code, and use feedback to repair their failed attempts >
Marcus: They tested five different ways the agent could call tools, including web search and Python programming with Morpheus's command line interface or even using a dedicated agent skill or an MCP server to connect all of these things >
Yuki: It’s important that they are testing those tool setups because how you set up the environment really determines if this whole autonomous loop actually works reliably across different biological problems, and that’s a key part of what they cover >
Ines: And when it comes to validation, the paper describes Morpheus separating the model from the simulator code so each simulation can be validated against a schema that keeps models small and limits the scope for errors >
Marcus: The agents only write model files in this standardized MorpheusML format, which means they don't get to alter the actual simulator code, which is a security requirement for wider use >
Yuki: So the paper is essentially showing that an AI agent can take a set of biological observations and turn them into a working virtual tissue model file without needing human intervention for every single step >
Conclusion: Ines: Looking at this paper "Morpheus.AI: agents autonomously generate, execute and refine virtual tissue models" by Prerana Chandratre and her team, the implication is that we might see a shift where the creation of these mechanistic models isn't just for specialists anymore >
Marcus: Yeah, it suggests that if we can get reliable agents to build these models from scratch using tools like Morpheus, then researchers in fields like cancer biology or developmental biology could start prototyping hypotheses much faster >
Yuki: What I think is that the real impact is democratizing the process; it means you don't need a PhD in advanced math just to start making those kinds of connections between cell migration and tissue morphology >
Ines: It’s about moving from slow, specialist work to something that could become routine, which is what they claim with their agents meeting eighty-three to eighty-six percent of study criteria in under two hours on average >
Marcus: And it points toward a future where the process of generating these complex biological models becomes much more accessible, provided you have a good model and the right large language model to guide the agent >
Yuki: So we're talking about making those kinds of connections between biology and mechanism more routine for a broader audience, which could accelerate discovery across many different biological domains >
Prerana Chandratre, Anjali Sharma, Justin Bürger Michaelis, Diego Jahn, Jörn StarrußsƟ, Δ Lutz Brusch, * and Simon Syga *
Center for Interdisciplinary Digital Sciences, TUD Dresden University of Technology
q-bio.QM, q-bio.CB, q-bio.TO
Submitted: 2026-10-29
Updated: 2026-10-29
Comments: 2 figures
Code: https://github.com/sisyga/morpheus-skills
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 91/100
The gist: Virtual tissue models link observations to mechanistic hypotheses, but building them demands expertise in biology, mathematics and programming <ref:2611.05413#pg2>.
Key concepts
- Morpheus.AI Agents
- These are specialized large language model agents designed to autonomously generate, run, and improve virtual tissue models. They work by calling tools and using feedback to correct their simulations until they meet specific study criteria.
- Virtual Tissue Models
- These are digital representations of biological tissues created through mathematical frameworks. Building them is complex, requiring experts in biology and math to translate observations into a structure that a simulator can run.
- Agentic Workflow
- This describes the process where the LLM agents actively interact with tools, write their own code, execute simulations, and use the results to self-correct. This loop allows them to build sophisticated models without constant human intervention.
Terminology
Summary
Virtual tissue models link observations to mechanistic hypotheses, but building them demands expertise in biology, mathematics and programming <ref:2611. p. 0 syga@tu-dresden. de)
Virtual tissue models link observations to mechanistic hypotheses, but building them demands
expertise in biology, mathematics and programming. de)
Virtual tissue models link observations to mechanistic hypotheses, but building them demands
expertise in biology, mathematics and programming. Here, large language model (LLM) agents
paired with the simulator Morpheus (Morpheus. Here, large language model (LLM) agents
paired with the simulator Morpheus (Morpheus. AI) autonomously built such models from fifteen
benchmark studies, five without a published model. AI) autonomously built such models from fifteen
benchmark studies, five without a published model. With a frontier LLM, agents that ran their
simulations met 83–86% of study-specific criteria in under two hours and for under US$4 per model. With a frontier LLM, agents that ran their
simulations met 83–86% of study-specific criteria in under two hours and for under US$4 per model. Such agents could make in silico modeling routine.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
The Agentic Workflow
The paper introduces 4 A modeler must
abstract observations into assumptions, cast them in a
mathematical framework, implement them in a simulator
such as PhysiCell,5 CompuCell3D6 or Morpheus,7 and
check the output against expected behavior; each step
needs different expertise (Fig. Large language model (LLM) agents call tools, run their
own code and use the feedback to repair failed attempts. AI” and gave them
five alternative tool setups that mimic common ways
of using LLMs (Fig. 1c,d; Methods): web search and
Python (Web), as in a chatbot; the Morpheus commandline interface (CLI), as in an agentic harness; the CLI with
only the documentation, no web search (Docs); an agent
skill including the reference library (Skill); or all of these
through an MCP server (MCP). No human intervened
while the agent generated, ran, inspected and revised its
models, and every attempt was kept for audit (Fig. We evaluated the agent on fifteen studies spanning cell
migration, morphogenesis, pattern formation, development, cancer and regeneration (Fig.Morpheus.AI
agents that autonomously generate, execute and refine virtual tissue models <ref:2611.05413#pg3>. These agents call tools, run their own code and use the feedback to repair failed attempts <ref:2611. p. 0Where the paper says it
Where the paper says it
Where the paper says it
Model Construction and Validation
Building a multiscale virtual tissue model is described as a slow, specialist work requiring abstraction into assumptions and casting them in a mathematical framework <ref:2611.05413#pg3>. Morpheus separates the model from the simulator code, allowing each simulation to be validated against a schema which keeps models small and limits the scope for errors <ref:2611. p. 0 Here, we pair LLM agents with the widely used modeling and simulation framework Morpheus,7 which suits
this test for four reasons. First, Morpheus separates the
model from the simulator code: each simulation is a single declarative MorpheusML file that can be validated
against a schema, which keeps models small, limits the
scope for errors and lets a modeler check what the agent
built without reviewing simulation code. Second, a curated library of validated reference models lets us test
how much examples help the agent. Third, the agent
writes only model files, which the trusted Morpheus simulator turns into simulations and runs, so agents cannot
alter the simulator code, a security requirement for wider
use.Where the paper says it
Where the paper says it
Performance and LLM Comparison
With GPT-6.1 Sol, every setup that could run Morpheus reached a mean score of 0. p. 0 With GPT-6. 1 Sol, every setup that could
run Morpheus reached a mean score of 0. 86, after
105 min to 110 min of active agent time and for US$3. 74
per run on average. 86),
suggesting that the agent formulated new mechanisms
about as well as it re-implemented published ones. To get there, the agent had to simulate its model and
inspect the output. The choice of LLM mattered more than the setup. 1 Sol outscored GPT-6 Luna in every setup (0. The choice of LLM mattered more than the setup. 1 Sol outscored GPT-6 Luna in every setup (0. Studies without a published model
scored about as high as those with one (0. 83 versus 0. 85
versus 0. 66 across the setups that could run Morpheus;
p ≤ 0. 006) at 10. 8 times the cost per run (geometric-mean
ratio, 95% CI 8.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Scoring and Future Directions
A separate LLM scorer agent checks study-specific qualitative and quantitative criteria against the agent-generated model and returns a score (0–1) <ref:2611.05413#pg3>. The automated score agreed exactly with manual expert scores in 139 of 225 runs, with Pearson r = 0. p. 5 5,
reasoning effort xhigh, three repetitions) by hand with
the same criteria. Automated and manual scores agreed
exactly in 139 of 225 runs (62%), with a mean absolute
difference of 0. 55, against 0. 86
for every setup that could run Morpheus; with GPT-6
arXiv:2611. Each model states
its mechanistic hypotheses explicitly in MorpheusML, so
they can be simulated, compared with alternatives and
refined in silico before a costly wet-lab experiment. Such
agents could make mechanistic modeling routine in biology, also for researchers without training in mathematics
and programming. The scorer is an LLM, and LLM
judges can favor answers like their own;22 ours (GPT6 Luna) is also one of the two agent LLMs, which is
why we validated it against manual scores. We tested
a single simulation platform and LLM provider, so the
benchmark tests whether the autonomous loop works and
which ingredients it needs. AI scientists that generate hypotheses, analyze data
or automate research end to end promise to accelerate
discovery,23–25 but audits of their trace logs and code have
revealed data leakage, metric misuse and post-hoc selection bias. 26 We therefore kept generation and evaluation
apart: the agent hands over only the model file in the
standardized MorpheusML format, which we re-execute in
Morpheus without the agent and score against criteria it
never saw, using an LLM scorer validated against human
expert modelers.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Methodological Details
The five setups differ only in the tools exposed to the agent; task, prompt, time budget and scoring were identical <ref:2611.05413#pg6>. The benchmark comprised 15 studies (Fig. p. 4 Benchmark
While an agent can be tasked with custom problem
statements and data, we standardized the task input
to the content of a single scientific paper. The benchmark comprised 15 studies (Fig. The remaining five (studies 11–15) report experimental observations without a computational
model, so the agent had to formulate the mechanism itself:
gastruloid axis formation,36 digit patterning,37 embryo
implantation,20 limb regeneration38 and liver organoid
zonation. 39 With its criteria, the benchmark can also
test other agentic modeling frameworks:11,17 even after
excluding the studies whose published model used a given
framework, more than ten problems remain new to it. 025 for each pair). The agent adapted models to what Morpheus can 32 for
GPT-6 Luna. Data availability
The data underlying Fig. Data availability
The data underlying Fig. 2 are available in the code
repository and its Zenodo archive (Code availability).Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.
Improvements for AI systems
-
Improve agent capability to handle model formulation for unpublished phenomena by enabling agents to
formulate new mechanisms
for studies without published models, as seen in Figure 2d iii. This allows agents todiscover and formulate mechanisms for studies without a published model,
making them capable of generating hypotheses where none exist. -
Enhance agent reasoning through superior LLM selection, as the study shows that GPT-6.1 Sol
outscored GPT-6 Luna in every setup
at 10.8 times the cost per run and achieved higher mean scores of 0.83–0.86 compared to Luna's 0.11 gap against a score of 0.60–0.73, suggesting that the choice of LLM is more critical than the setup for performance. -
Strengthen agent robustness by optimizing tool usage based on LLM sensitivity, as GPT-6 Luna was
sensitive to its tools,
scoring lowest without web search (Docs; 0.60) but best with the plain CLI and web access (0.73), indicating that a more sophisticated tool set, like Setup 5 (MCP), is necessary for optimal performance across different LLMs. -
Implement a decoupled generation and evaluation pipeline to prevent bias, as the authors state they
kept generation and evaluation apart: none of the agent’s own simulation outputs reached the scorer,
ensuring thatgeneration and evaluation were kept separate.
This separation validates that the scoring agent judges only against criteria it has never seen, mitigating potential self-favoring biases. -
Increase model fidelity by leveraging structured validation frameworks, as Morpheus's declarative XML format
can be validated against a schema, which keeps models small, limits the scope for errors and lets a modeler check what the agent built without reviewing simulation code.
This structural constraint helps ensure that agents build models that are less prone to unsimulatable or erroneous constructs.
Abstract
Virtual tissue models link observations to mechanistic hypotheses, but building them demands expertise in biology, mathematics and programming. Here, large language model (LLM) agents paired with the simulator Morpheus (Morpheus.AI) autonomously built such models from fifteen benchmark studies, five without a published model. With a frontier LLM, agents that ran their simulations met 83-86% of study-specific criteria in under two hours and for under US4 per model. Such agents could make in silico modeling routine.
Sources
Related papers
- ViroGym: Realistic Large-Scale Benchmarks for Evaluating Viral Proteins
- CellCanon: An Agentic Cellular World Model for Prediction, Explanation, and Observation-Driven Updating
- Probing and steering biology across Boltz-1s trunk-diffusion boundary
- Trajectory Inference of Human Aging from Cross-Sectional DNA Methylation Data
- Bridging the Modality Bottleneck in Pathology MIL through Molecularly Guided Feature Adaptation
- Molecular notation can act as a shortcut in AI benchmark evaluation