Until recently, most people viewed AI in biomedical research as a support tool. It could summarize papers, explain genes, help write Python scripts, generate hypotheses, or assist with literature review.
But STELLA represents a more advanced direction.
It is not just a chatbot that answers biomedical questions. It is designed as a self-evolving multimodal biomedical agent system that can reason, create tools, run computational workflows, evaluate its own results, learn from feedback, and connect predictions with experimental validation.
In simple terms, STELLA is trying to move from being an AI assistant to becoming an AI research partner.
And this is a very important shift.
Why traditional AI systems are not enough for biomedical research
Biomedical research is complex because it is not a single-step reasoning problem.
A real research question may require:
reading scientific literature
querying biomedical databases
analyzing gene expression data
understanding protein structures
checking pathway and disease associations
designing mutations
running simulations
generating code
interpreting biological results
validating hypotheses in the wet lab
A normal LLM may understand the language around these tasks, but it often lacks three important capabilities:
It cannot reliably execute the full scientific workflow.
It depends on a fixed set of tools or static knowledge.
It does not improve its research strategy through experimental feedback.
This is where agentic AI becomes interesting.
Instead of treating AI as a single model that gives one answer, agentic systems break a complex problem into multiple roles, tools, actions, validations, and feedback loops.
That is the core idea behind STELLA.
The agentic architecture behind STELLA
Conceptual view of how STELLA interacts with the biomedical research environment.
STELLA uses a collaborative multi-agent design. Each agent has a specific role in the research pipeline.
1. Manager Agent
The Manager Agent acts like the research planner. It receives a high-level biomedical goal and converts it into a structured research workflow.
For example, instead of simply answering:
“Find a novel target for NK-cell therapy in AML”
the Manager Agent can break the problem into steps such as:
define the disease and phenotype context
identify candidate genes or proteins
search relevant databases
analyze pathway and network evidence
filter already-known targets
rank novel candidates
prepare a biological rationale
This is similar to how a senior scientist or research lead would design an investigation.
2. Developer Agent
The Developer Agent is the execution layer.
It writes and runs code, usually Python-based workflows, to perform actual bioinformatics analysis.
This is important because serious biomedical research cannot rely only on natural language reasoning. It needs executable pipelines.
The Developer Agent can work with datasets, scripts, APIs, model outputs, and analysis tools. It turns the research plan into computational action.
In software engineering terms, this is where the agent moves from “thinking” to “doing.”
3. Critic Agent
The Critic Agent evaluates intermediate outputs. This is one of the most important parts of the architecture.
In real research, the first result is often incomplete, weak, or biologically unconvincing. A good scientist does not stop at the first answer. They ask:
Is this result actionable?
Is the biological logic strong enough?
Are we only describing correlation?
Do we need mechanistic evidence?
Do we need another database, model, or experiment?
Are we missing a better explanation?
The Critic Agent plays this role. It reviews results, identifies gaps, and pushes the system to refine the workflow. This makes the pipeline more robust than a one-shot LLM answer.
4. Tool Creation Agent
The Tool Creation Agent is what makes STELLA especially interesting.
Most AI agents depend on prebuilt tools. But biomedical research constantly changes. New databases, new APIs, new models, and new experimental methods appear all the time.
A fixed toolkit becomes outdated quickly.
STELLA addresses this through a dynamic Tool Ocean.
When the system identifies that it lacks a required capability, the Tool Creation Agent can discover, wrap, validate, and integrate new tools.
This may include tools for:
disease-gene association queries
CRISPR or RNAi dependency analysis
protein-DNA interaction analysis
restriction enzyme analysis
protein structure prediction
single-cell data interpretation
pathway and network analysis
custom biomedical database access
This is a key idea: the agent does not only use tools; it can expand its own tool ecosystem.
That is a big step toward self-evolving AI systems.
Template Library: how the agent learns from successful workflows
Another powerful concept in STELLA is the Template Library. When a research workflow succeeds, STELLA can store the reasoning pattern as a reusable template.
For example, if a workflow successfully moves from:
disease context → candidate genes → pathway evidence → patient dataset validation → target ranking
then this strategy can be reused for future target discovery tasks.
This is similar to how experienced researchers build mental models over time. A beginner sees each task as new. An expert recognizes patterns.
STELLA attempts to capture these successful patterns and reuse them.
This means the system is not only executing tasks. It is gradually improving its own research process.
Agentic pipeline: from question to scientific hypothesis
The most interesting part of STELLA is the full agentic pipeline.
A simplified version looks like this:
User goal → Manager planning → Tool selection → Code execution → Result evaluation → Tool creation if needed → Re-analysis → Hypothesis generation → Wet-lab validation → Feedback into the system
This is very different from a normal chatbot workflow.
A chatbot usually does:
Question → Answer
An agentic biomedical system does:
Question → Plan → Tools → Execution → Critique → Iteration → Evidence → Validation → Learning
That is closer to how real scientific discovery works.
Example 1: Discovering a new NK-cell target in AML
One of the paper’s strongest examples is in cancer immunology.
STELLA was asked to identify novel, previously unreported negative regulators of NK-cell function in the context of acute myeloid leukemia, or AML.
NK cells are important immune cells that can kill cancer cells. But in AML, their activity can be suppressed by inhibitory signals from tumor cells and the microenvironment.
STELLA used its reasoning templates and biomedical tools to search literature, databases, regulatory networks, and patient evidence.
It then ranked candidate genes and identified BTN3A1 as a promising negative regulator of NK-cell function.
The important part is that this was not just a theoretical prediction.
The researchers validated it experimentally.
They used CRISPR knockout in NK92MI cells and tested the effect against multiple AML cell lines. The result was that loss of BTN3A1 increased NK-cell cytotoxicity and degranulation.
This is a strong example of an agentic AI system supporting the complete research loop:
hypothesis generation → candidate prioritization → experimental validation
Example 2: Enzyme engineering with iterative feedback
The second important example is protein engineering.
STELLA was used to optimize strictosidine synthase, an enzyme involved in biosynthesis of terpenoid indole alkaloids, which are connected to medically valuable compounds.
The agentic pipeline here involved:
retrieving enzyme and substrate information
predicting enzyme-substrate structures
identifying catalytic pocket residues
generating a virtual mutant library
predicting activity using a protein language model
filtering variants using stability calculations
recommending mutations for wet-lab testing
using experimental feedback to improve the next round
This is very close to a closed-loop engineering system.
Instead of manually testing random mutations, the agent guides the search through computational prediction, structural reasoning, and experimental feedback.
One highlighted variant, M276L, showed more than a two-fold improvement in catalytic activity.
This is important because it shows that agentic AI can be useful not only for analyzing biology, but also for designing better biological systems.
Example 3: From computational intelligence to physical lab intelligence
The third part of STELLA moves into robotics.
This is where the idea of a “Biomedical World Model” becomes more ambitious.
The system works with Vision-Language-Action models, or VLA models, for biomedical lab tasks.
These tasks are not easy.
A robot in a lab needs to handle:
pipettes
centrifuge tubes
digital panels
transparent containers
liquid levels
rotating knobs
tube caps
precise placement tasks
STELLA uses a Decompose-Monitor-Recover mechanism.
This means the task is broken into smaller subtasks.
For example, screwing a centrifuge tube cap may be decomposed into:
grasp the tube
align the cap
engage the thread
tighten the cap
During execution, the system monitors whether each step is going correctly. If a failure happens, it uses recovery tools to correct the mistake.
Those recovery attempts then become training data for improving the VLA model. This is a powerful idea. The system learns not only from successful demonstrations, but also from failures and recoveries. That is closer to how humans learn physical skills.
Why this matters for engineers
As a software engineer learning bioinformatics, I find this direction very exciting.
Because the future of biomedical AI will not be only about biology knowledge.
It will need engineering skills such as:
data pipelines
API integration
distributed systems
agent orchestration
workflow automation
Python-based scientific computing
database querying
model evaluation
cloud-native infrastructure
reproducible research environments
human-in-the-loop validation
In many ways, agentic biomedical systems look like advanced software platforms.
They combine AI models, tools, data, execution environments, feedback loops, and observability.
For engineers, this creates a huge opportunity.
The next generation of biomedical platforms may need people who understand both:
software architecture and scientific workflows
The bigger picture: AI systems that evolve with science
The most important takeaway from STELLA is not that one model performed well on a benchmark.
The bigger takeaway is this:
The future may belong to AI systems that can continuously evolve with the scientific environment.
A static AI model is limited by its training data.
But a self-evolving agent can:
learn from successful workflows
create new tools
call external databases
execute code
critique its own outputs
incorporate experimental feedback
improve future decision-making
This makes the system more adaptable.
And adaptability is critical in biomedicine, because biology is complex, noisy, and constantly changing.
But we should be careful
Biomedical AI also requires strong human oversight.
A system that proposes targets, mutations, or experimental strategies must be evaluated carefully. Wrong conclusions in biomedical research can waste resources or even create safety risks if moved too quickly toward clinical use.
So the right framing is not:
“AI will replace scientists.”
The better framing is:
“Agentic AI may become a powerful research partner for scientists.”
My takeaway
STELLA gives us a glimpse of how AI could transform biomedical research.
Not as a simple question-answering system.
But as a self-evolving research engine that can:
plan scientific workflows
create and use tools
execute bioinformatics pipelines
evaluate intermediate results
generate hypotheses
connect computation with wet-lab validation
improve from feedback
support robotic lab execution
This is where AI becomes much more than a chatbot. It becomes an agentic system capable of participating in the scientific discovery process.
For someone coming from software engineering and exploring bioinformatics, this is extremely inspiring. It shows that the future of life sciences will need people who can build bridges between AI, biology, data engineering, automation, and real-world experimentation.
The next wave of impactful software may not only power apps and dashboards. It may power discovery pipelines, autonomous labs, and new biological insights.