Zero-Trust Edge Architecture
Molecular data is rarely just a file. A structure can carry a target hypothesis, an assay condition, or an entire programme direction. Our edge-native architecture starts by treating that context as part of the data boundary.
The boundary is part of the workflow
A practical workflow has to move structures through standardisation, feature preparation, inference, ranking, and review. Sending each step to a hosted service can make ownership and provenance difficult to reason about.
The first boundary is chemical, not just network-level. A SMILES string may need sanitisation, charge normalisation, stereochemistry checks, and a decision about whether explicit hydrogens or a 3D conformer are required. Those transformations should be recorded instead of silently discarded.
The Moloracle direction is local-first: keep molecular inputs and intermediate artefacts on hardware the team controls, then make each model run explicit about what entered the system and what came out.
Zero trust is operational, not a slogan
The useful question is not whether a system sounds private. It is whether access, model versions, generated artefacts, and review decisions can be inspected and limited in practice.
That means least-privilege access, versioned inputs, clear export boundaries, and an evidence trail that follows a candidate from prediction to validation. The security posture still depends on the final deployment and its operational controls.
Local execution is not automatically secure. A workstation can still be misconfigured, a model artefact can still be copied, and a service account can still have excessive permissions. The useful design goal is to make those risks visible and controllable rather than implying that edge is a substitute for security engineering.
What changes for a model run
A model output becomes more useful when it arrives with the structure identifier, preprocessing choice, model version, uncertainty, and next-step recommendation. This is the foundation for a reviewable result rather than an opaque score.
A reproducible record can include a hash of the input file, the normalised representation, the conformer-generation settings, the runtime version, and the output artefact. It does not make a prediction correct, but it makes a later disagreement diagnosable.