Caelis Neural Base 1
New computational results from CNB-1: from the first structured biological design to a second defined protein candidate, AGD-F1.
Abstract
Caelis Neural Base-1 (CNB-1) is the first experimental model developed within Project Proteus. Its early experiments were designed to answer a foundational question: can a protein-design model produce defined biological sequences that can be converted into concrete molecular candidates for computational investigation?
The answer so far is a documented set of structured outputs. The first major experiment produced an approximately 350-amino-acid hydrolytic enzyme candidate targeting PET degradation. A subsequent CNB-1 generation produced AGD-F1, a 377-amino-acid computationally designed protein candidate generated from a structural design seed.
These results do not establish biological activity. They establish that CNB-1 has produced defined candidates that can enter a deeper sequence, structure and validation pipeline.
1. CNB-1 and Proteus
Project Proteus is the research program. Caelis Neural Base-1 is an individual experimental model within that program. CNB-1 is therefore not the name of Proteus itself; it is the first model used to investigate AI-assisted protein design within the project.
PROJECT Project Proteus MODEL Caelis Neural Base-1 ABBREVIATION CNB-1 ORGANIZATION Inserloft Research ROLE First experimental Proteus model OUTPUT Computational protein candidates STATUS Experimental research
2. The first documented design
CNB-1’s first documented protein-design experiment produced an approximately 350-amino-acid candidate designed around a hydrolytic-enzyme objective.
The sequence was initiated using the MNPAQ structural seed, associated in the project documentation with a signal-peptide motif found in natural PETases from Ideonella sakaiensis. The target hypothesis was PET hydrolysis into terephthalic acid and ethylene glycol.
The candidate was subsequently treated as a computational design hypothesis. Proposed features such as a Ser–His–Asp catalytic triad and an alpha/beta-hydrolase architecture remain validation targets rather than experimentally established properties.
3. A new CNB-1 output: AGD-F1
CNB-1 has now produced another defined candidate: AGD-F1, a 377-amino-acid computationally designed protein candidate.
AGD-F1 was generated using CNB-1 under a structural-seed-conditioned protein sequence generation workflow. Its resulting sequence was subsequently evaluated with ESMFold to obtain a computational structural hypothesis.
The designation AGD-F1 refers to the prominent representation of alanine, glycine and aspartate within the candidate sequence and its resulting sequence-level characteristics. It does not imply a demonstrated biochemical function.
4. What AGD-F1 tells us
AGD-F1 is significant as another concrete output of the CNB-1 generation pipeline. The result is not simply an arbitrary text sequence: it is a defined 377-residue molecular candidate that can be represented as a structural model and subjected to further analysis.
Its current computational record includes sequence characterization, an initial sequence-similarity analysis and an ESMFold structural prediction. The initial similarity analysis did not identify a clear high-confidence match to a known protein family in the reference analysis used for the candidate.
That observation does not establish evolutionary novelty, a novel fold or a new biological function. Profile-based homology analysis and structure-based searches remain necessary.
5. Structure is still a hypothesis
The AGD-F1 structure currently available through the Inserloft Data Protein Bank is a computational prediction. It is not an experimentally determined structure.
Further evaluation should examine per-residue confidence, global confidence, secondary structure, backbone geometry, contacts, hydrophobic-core organization, surface exposure, compactness and similarity to experimentally characterized structures.
Only experimental characterization can establish whether the predicted conformation corresponds to a stable physical protein and whether the candidate has measurable biochemical activity.
6. The CNB-1 evidence chain
The current Proteus workflow is deliberately staged:
- Generation — CNB-1 produces a defined amino-acid sequence from a computational design condition.
- Sequence analysis — the candidate is inspected for composition, constraints and similarity.
- Structure prediction — a model such as ESMFold provides a structural hypothesis.
- Structural characterization — fold, confidence, geometry and molecular organization are examined.
- Functional analysis — candidate-specific mechanistic hypotheses are tested computationally.
- Experimental validation — laboratory measurements are required before biological claims can be made.
CNB-1 is currently producing candidates that can move through the early stages of this chain. The work ahead is to determine how many survive increasingly demanding tests.
7. What comes next
For AGD-F1, the next stage is deeper computational characterization: profile-based homology searches, structure-based searches, independent structure prediction, structural confidence analysis, fold classification, stability estimation and surface or pocket characterization.
For the PET-hydrolase candidate, future work can further examine the proposed catalytic architecture and the structural requirements of the PET-hydrolysis hypothesis before any experimental claim is made.
At the model level, the objective is to repeat the generation-and-evaluation loop across additional candidates and model versions, making it possible to distinguish genuine improvements from isolated successful generations.
10. Conclusion
Two days after the first documented CNB-1 experiments, the model has produced another concrete protein-design candidate: AGD-F1.
That does not mean Proteus has already discovered a functional protein. It means the system is producing molecular hypotheses that can be inspected, modeled, challenged and tested.
Generate. Structure. Analyze. Validate. Repeat.
Inserloft Research · Project Proteus · Caelis Neural Base-1 · Technical Article 001