Exploring the digital sandboxes where biological discoveries are born
Imagine if scientists could test life-saving drugs not in lab animals or human patients, but in intricate digital universes designed to mimic our own biology. This isn't science fiction—it's the cutting edge of computational biology, where researchers create credible world models to simulate everything from cellular processes to entire ecosystems. These models serve as digital sandboxes where biological theories can be tested, refined, and validated before a single test tube is lifted in a laboratory.
The concept of credible worlds comes from economics and philosophy 4 , but it has found an extraordinarily productive home in biological sciences. As we face increasingly complex challenges—from personalized cancer treatments to ecosystem conservation—these models provide something previously unimaginable: a way to explore countless biological scenarios rapidly, inexpensively, and ethically. Recent advances have made these simulations so accurate that they can outperform human experts in certain tasks 1 , while also raising important questions about how we validate these digital realities.
Credible world models are computational simulations that researchers construct to represent biological systems with sufficient realism to draw meaningful conclusions about how those systems operate in nature. Unlike simpler models that might focus solely on predicting outcomes, credible worlds aim to recreate the underlying mechanisms that drive biological phenomena.
Philosophically, these models operate on the principle that if you can build a working simulation that produces results matching real-world observations, you likely understand the key mechanisms involved—a notion echoing physicist Richard Feynman's famous statement, "What I cannot create, I do not understand" 8 .
The model's internal workings must reflect biologically plausible processes, even if simplified.
The model should generate complex phenomena from simple rules, mimicking how real biological systems operate.
The model's outputs must align with experimental data from laboratory studies or natural observations.
Traditional biological models largely relied on mathematical equations to describe relationships between variables. While useful, these equation-based models often struggled to capture the complexity of biological systems where countless elements interact simultaneously.
The emerging synthetic modeling approach addresses this limitation by creating executable analogues—computational systems that mimic biological processes through step-by-step operations rather than mathematical summary 8 . These models more closely resemble their wet-lab counterparts than traditional computational models do, functioning as "virtual laboratories" where researchers can observe biological mechanisms in action.
One of the most powerful aspects of credible world modeling is its ability to facilitate abductive reasoning—the process of forming hypotheses about what might possibly be true 8 . When researchers observe unexpected phenomena emerging from their models, they can generate novel hypotheses about how similar phenomena might arise in actual biological systems.
As one researcher notes, synthetic models "encourage and facilitate abductive reasoning, a primary means of knowledge creation and creative cognition" 8 . This ability to generate entirely new insights makes credible world modeling particularly valuable for exploring biological questions where existing knowledge is limited.
In a remarkable demonstration of credible world modeling's potential, researchers recently tested whether artificial intelligence could outperform human experts in virology problem-solving 1 . The study, conducted by researchers from the Center for AI Safety, MIT's Media Lab, and other institutions, created an extremely difficult practical test measuring the ability to troubleshoot complex lab procedures and protocols.
The research team consulted virologists to develop questions that represented non-Google-able challenges—the kind of practical knowledge that is typically passed down from experienced colleagues rather than found in academic papers. These questions took the form: "I have been culturing this particular virus in this cell type, in these specific conditions, for this amount of time. I have this amount of information about what's gone wrong. Can you tell me what is the most likely problem?" 1
The researchers administered these practical virology problems to both PhD-level virologists and various AI models, including OpenAI's o3, Google's Gemini 2.5 Pro, and Anthropic's Claude 3.5 Sonnet. The performance was measured based on accuracy in identifying the correct solutions to these complex troubleshooting scenarios.
| Solver Type | Specific Model | Accuracy Score (%) |
|---|---|---|
| Human Expert | PhD-level Virologists | 22.1 |
| AI Model | OpenAI o3 | 43.8 |
| AI Model | Google Gemini 2.5 Pro | 37.6 |
| AI Model | Claude 3.5 Sonnet (Jun '24) | 26.9 |
| AI Model | Claude 3.5 Sonnet (Oct '24) | 33.6 |
The results were striking: every AI model outperformed human experts, even within the virologists' own areas of specialization 1 . Moreover, the AI models showed rapid improvement over time—Claude's accuracy jumped nearly 7 percentage points in just four months.
This experiment demonstrates how AI systems can develop practical biological knowledge without direct laboratory experience, essentially learning from the collective knowledge embedded in their training data. The implications are dual-edged: such systems could dramatically accelerate biomedical research but also raise concerns about potential misuse if fallen into wrong hands 1 .
Creating credible world models requires both conceptual frameworks and practical tools. Here are the key components researchers use to develop and validate these biological simulations:
| Tool Category | Specific Examples | Function |
|---|---|---|
| Modeling Languages | SBML (Systems Biology Markup Language) | Standardized format for encoding mathematical models of biological processes |
| CellML | XML-based language for reproducing mathematical models of any kind | |
| BioPAX | Ontology for describing biological pathway data | |
| Annotation Standards | MIRIAM Guidelines | Minimum information requirements for model annotation |
| SBMate | Python package for assessing semantic annotations in models | |
| Simulation Approaches | Agent-Based Models | Simulating actions of individual entities to observe emergent system behavior |
| Discrete Event Simulation | Modeling system operation as a sequence of events in time | |
| Ordinary Differential Equations | Describing systems through rates of change of continuous variables |
For credible world models to be scientifically useful, they must be reproducible and built upon shared standards. The systems biology community has developed extensive standards for model encoding, annotation, simulation, and dissemination 2 . These include:
The most widely used model format, supported by over 200 third-party tools 2
Providing minimum information requirements for model annotation 2
Ensuring models can be recreated and verified by independent researchers
Creating a biological model is one thing; establishing its credibility is another. Regulatory agencies like the FDA and EMA have begun developing standards for assessing model credibility when considering computational evidence for pharmaceutical and medical device approval 2 . The FDA defines model credibility as "the trust, established through the collection of evidence, in the predictive capability of a computational model for a context of use" 2 .
Researchers use several strategies to validate their credible world models:
Checking whether model outputs match experimental data
Testing whether models can correctly predict future observations
Assessing whether the model's internal processes align with biological knowledge
As models become more sophisticated, validation approaches continue to evolve, incorporating new statistical methods and comparative frameworks.
The future of credible world modeling in biology involves integration with several cutting-edge technologies:
Providing unprecedented resolution for modeling biological systems 6
Offering improved validation platforms for model predictions 7
potentially solving currently intractable computational challenges in simulation
As modeling capabilities advance, they raise important ethical questions. The same AI systems that can outperform virologists in troubleshooting 1 could potentially be misused to design biological threats. Researchers and policymakers are increasingly focused on developing safeguards—including gated access to powerful models and ethical guidelines for their use 1 .
| Potential Benefits | Associated Risks | Proposed Safeguards |
|---|---|---|
| Accelerated drug discovery | Misuse for bioweapon development | Gated access for trusted researchers |
| Democratized biological expertise | Reduced barriers to dangerous knowledge | Input and output filters on AI systems |
| Improved pandemic preparedness | Unanticipated consequences | Red-team testing and risk assessment |
| Personalized medicine applications | Privacy concerns with patient data | Anonymization and data protection protocols |
Credible world modeling represents a fundamental shift in how we do biological science. By creating executable analogues of biological systems, researchers can explore questions that would be impossible, impractical, or unethical to investigate in traditional laboratory settings. These digital sandboxes have evolved from simple mathematical representations to sophisticated simulations that can outperform human experts in specific domains 1 .
As the technology advances, the line between digital and physical biology continues to blur. The recent demonstration of AI systems surpassing human virologists in practical problem-solving 1 hints at a future where biological discovery happens increasingly through interaction with credible world models. This doesn't eliminate the need for traditional laboratory work but rather complements it, allowing researchers to explore possibilities more efficiently before validating findings in physical experiments.
The challenge ahead lies not only in improving the technical capabilities of these models but also in developing appropriate safeguards, ethical guidelines, and validation standards. If we can navigate these challenges responsibly, credible world modeling promises to accelerate our understanding of biology and revolutionize how we approach human health and disease treatment.
As one researcher aptly notes, "When synthetic models are executed, we observe different aspects of knowledge in action from different perspectives" 8 . These digital credible worlds have become more than just tools—they're becoming partners in the scientific process, expanding our capacity to understand and manipulate the biological world in ways previously confined to the realm of science fiction.
References will be added here in the final publication.