03Technology

One stack: L++ → Lifabrica → Life Code

The three-layer stack and the AI-plus-expert workflow behind every model — technical enough for a scientific evaluator, plain enough for a non-coder.

—The stack

Lifabrica sits on L++, which sits on Life Code

product · Oct 2026

Lifabrica

Our flagship: the hosted, conversational way to reach every model Omphalos and its community build. Ask a natural-language question, get a simulation-backed answer — not an LLM guess. Domain scientists run experiments, apply perturbations, and inspect results without writing code; the model source stays available underneath for anyone who wants to look directly at the mechanism.

engine

L++

The biology-native language and simulation engine Lifabrica runs on — the backend that makes it possible for an answer to be a simulation rather than a guess. Learnable by a biologist in under a day; native syntax for chemistry, units, space, and time. Fully kinetic, multi-scale, and genetically or chemically perturbable.

commons

Life Code

An open, community-governed library of models and data, stewarded by the independent Life Code Foundation, and the growing library Lifabrica draws on. Contributors are credited and compensated for models and data that improve the commons, tracked against the code that actually runs.

—Method

How a model gets built

AI provides the hands and tools; the human expert is the control tower responsible for scientific judgment.

Draft

Generative AI assembles model components — reaction networks, parameter sets, compartmental structure — from literature and existing open-source models.

Inspect

A domain expert reviews the logic, assumptions, and parameters directly in L++ before anything is accepted. This is genuine authorship.

Simulate

The model runs. Plausible-sounding but structurally wrong reactions tend to surface here, visible in the output.

Calibrate & validate

The model is tested against experimental or omics data it was never fitted to, and revised.

Deploy

The validated model becomes reusable infrastructure — ported, extended, or deployed into a new application.

—Inputs

What the platform can ingest

Heterogeneous data

In vivo imaging, in vitro assay output, omics, serum biomarkers — integrated at the model level rather than pre-processed into one rigid format.

Incomplete data

Missing parameters start from literature priors and refine as data arrives. A model calibrated to 2 of 5 target organs still returns valid predictions for those 2 — no all-or-nothing dependency.

Existing published models

Any ODE or SBML model from the literature can be adopted into L++ in minutes, so new work starts from the current state of the art.

Reference-and-derive construction

A well-characterized reference organism is built first; strain- and species-specific models are derived from it by encoding known mechanistic differences — modular and reusable, the way software is built.

Data posture

Not trained on your data

There is no training corpus behind an L++ model. Inputs are knowledge — biochemistry, enzyme kinetics, published regulatory-network structure. Multi-omics data is used to validate a model after it’s built, never to fit it. That makes an external, independent dataset genuinely independent evidence, with no risk of the overlap a shared benchmark can create.