Project
Proteus
AI for the next generation of biological discovery.
Proteus explores how machine intelligence can help researchers understand, model and design biological systems — from proteins and molecules to broader biological discovery — turning computation into a tool for scientific progress.
. : + * # @ @ # * + : .
: + * # @@@# * + :
+ * # @@@@@# * +
* # @@@@@@@# *
# @@@@@@@@ #
Introducing Project Proteus
Discovery is a search problem.
Biology is not a simple system with a finite list of answers. It is a vast space of structures, sequences, interactions, behaviors and possible interventions. The challenge is not only understanding what is already known. It is deciding what deserves to be explored next.
Proteus approaches that challenge as a search problem: use computation to map possibility, reduce uncertainty and help researchers spend more of their time on the questions that can move science forward.
It should help determine what is worth knowing next.
The goal is not to predict everything. It is to make the next experiment more informative.
Intelligence should expand scientific possibility.
Proteus is built around a simple philosophy: AI should not replace scientific reasoning. It should extend it.
The role of an intelligent system is to help researchers see more possibilities, connect information that is difficult to connect manually, test hypotheses computationally, identify promising directions and make iteration faster. The final standard is not how impressive a model looks, but whether it helps produce better science.
That means staying curious without becoming careless, moving quickly without confusing speed for evidence, and treating every result as something to test rather than something to believe automatically.
Evidence over appearance
A useful system must be evaluated by reproducible results, meaningful benchmarks and scientific validation — not by demos alone.
Exploration before certainty
Early research should make room for unknowns. Proteus is intended to explore difficult spaces rather than pretend that the answers are already known.
Models serve the research
No single architecture is sacred. Models, datasets and tooling should change when experiments show a better path.
Intelligence meets biology.
Proteus is not defined by one model. It is a research program spanning multiple connected directions, with each line contributing a different piece of the discovery process.
Protein intelligence
Systems for learning from protein sequences, structures and functions, with the longer-term possibility of using those representations for prediction and design.
Generative biology
Investigating whether generative models can propose useful biological candidates and navigate design spaces rather than only describe existing observations.
Discovery systems
Connecting models, data, evaluation and scientific workflows into systems that support repeated cycles of hypothesis, computation and experimentation.
Biological representation
Exploring representations that allow very different biological information — sequence, structure, interaction and context — to become computationally useful together.
Candidate search
Building approaches for navigating large numbers of possible candidates while preserving useful scientific constraints and evaluation criteria.
Research infrastructure
Developing the data pipelines, evaluation systems, experiments and compute foundations required to turn research ideas into repeatable systems.
CaelisNeuralBase 3 is now part of the Proteus foundation.
Inserloft Research has released CaelisNeuralBase 3 (CNB 3), a public protein foundation model designed to give researchers a starting point for protein design, discovery and biological research.
For a better world. For a better life.
The purpose of Proteus is broader than a single disease, therapeutic category or biological target. The project is interested in the underlying technology required to make difficult biological discovery more tractable.
Think. Generate. Evaluate. Learn.
Proteus is envisioned as a continuous research loop rather than a one-shot prediction engine. A model produces a hypothesis or candidate, the system evaluates it, the result becomes information, and that information guides the next iteration.
Observe
Collect and organize biological information, experimental results and relevant scientific context.
Represent
Transform complex biological information into computational representations that models can work with.
Generate
Produce predictions, hypotheses or candidate designs from the available evidence and constraints.
Evaluate
Measure outputs against scientific criteria, simulations, benchmarks or experimental evidence.
Learn
Feed the results back into the system so that future iterations can improve instead of repeating the same search blindly.
Iterate
Repeat the cycle until the research question becomes clearer, the candidate space becomes smaller or the next experiment becomes more informative.
A research stack, not a single model.
Proteus is an evolving system of models, data, evaluation, scientific tooling and infrastructure working together. The aim is not to create one model that claims to solve biology, but to build a stack that helps make difficult scientific questions more tractable.
> loading biological representation layer
> connecting model + data + evaluation
> search space: biological systems
> objective: accelerate discovery
> status: ACTIVE
> awaiting next experiment
Fast research still needs slow truth.
AI can make it cheap to generate thousands of ideas. That does not make those ideas correct. Proteus therefore treats evaluation as part of the architecture, not as a final checkbox.
The project should continuously distinguish between what a model predicts, what a simulation suggests, what an experiment demonstrates and what remains unknown. A candidate can be interesting without being validated. A model can be useful without being right about everything. Research becomes stronger when those distinctions remain explicit.
Measure
Use meaningful evaluations rather than relying on subjective impressions of model quality.
Compare
Benchmark new approaches against credible baselines and understand where improvements actually come from.
Validate
Whenever possible, connect computational results to stronger forms of evidence, including appropriate experimental validation.
Capability without care is not progress.
Biology is powerful. Systems that operate in biological design spaces must therefore be developed with scientific responsibility, careful evaluation and awareness of misuse and unintended consequences.
Proteus is intended to prioritize legitimate scientific discovery and responsible research. As capabilities mature, the project should make decisions about publication, access, evaluation and deployment according to the risks and evidence associated with each system.
Scientific honesty
Do not present preliminary results as established science, and do not hide important limitations behind polished demonstrations.
Responsible access
Consider how research outputs can be shared in ways that maximize legitimate scientific value while reducing avoidable misuse.
Human judgment
AI should remain a research instrument. Important scientific and real-world decisions require appropriate expert judgment and evidence.
Research should leave a trail.
Proteus is part of Inserloft Research, where models and experiments are developed as an evolving body of work. The project should become more transparent as the underlying research becomes more mature and reproducible.
That means publishing technical details, experiments, benchmarks, limitations and results when they are ready — while avoiding the temptation to publish something merely because it sounds impressive.
Build. Test. Measure. Repeat.
Proteus is intentionally an unfinished system. Its architecture, models and research directions are expected to change as experiments reveal better questions and better methods.
The operating principle is simple: make it work, understand why it works, measure where it fails, improve it, and repeat. The project is not trying to jump directly to a mythical “AI scientist.” It is trying to build the pieces that could make increasingly capable scientific systems possible.
What comes next may not look like today’s models.
The long-term question behind Proteus is larger than any individual model release: what would it take for AI systems to become genuinely useful partners in difficult biological research?
Answering that may require better representations, stronger reasoning, simulation, retrieval, generative systems, scientific memory, evaluation loops, specialized infrastructure and eventually tighter connections between computation and experiments. No single component is enough by itself.
Proteus is therefore less a finished product than a direction: toward systems that can help researchers search biological possibility with greater speed, depth and precision while remaining grounded in evidence.
Caelis Neural Base 1.
CNB-1 is the first experimental model developed for Project Proteus. Its first documented tests produced structured biological outputs and an approximately 350-amino-acid candidate under a PET-hydrolase design objective.
Structured outputs
The first experiments produced interpretable biological sequence candidates rather than being treated as arbitrary strings.
Biologically informed seed
The documented PET-hydrolase experiment was initiated from the MNPAQ seed associated with a signal-peptide motif found in natural PETases from Ideonella sakaiensis.
Computational benchmark
CNB-1 has a current benchmark snapshot across five protein-model evaluation dimensions. The benchmark is evidence about measured tasks, not proof of biochemical activity.