Omnix · Predictive modelling
A predictive model that arrives with its evidence, not just a number
Omnix trains the model, then hands back a model card: the choices grounded in evidence, the evaluation audited for leakage (group-aware cross-validation), SHAP interpretability mandatory, and the research-only boundary enforced in the type system. Omnix predicts; the human decides.
trains a model and reports accuracy.
detects known resistance genes.
grounds the choices, audits for leakage, ships a model card, enforces the research-only boundary.
Model
Validation & tuning
Grounding & evidence
Preprocessing
Metadata
Threads the closed-loop lineage + literature grounding.
Data
View / edit raw data
Understanding Omnix
Omnix trains a predictive model and hands you back a model card, not just a number, with the model and feature choices grounded in evidence, the evaluation audited for leakage, the interpretability mandatory, and the research-only boundary enforced in the type system. It predicts; the human decides.
What Omnix is
Omnix is the predictive-modelling layer of the closed loop. It trains classification, regression, ranking, survival, and antimicrobial-resistance models on your analytical data, and produces a complete model card: performance, calibration, feature attribution, and the evidence behind every choice.
Predicting resistance, or any phenotype, from features is a mature, crowded field, and we do not claim to beat it on raw accuracy. Omnix is structurally different: every model arrives literature-grounded, leakage-audited, model-carded, and provenance-tracked, as a research instrument, never a diagnostic.
Why it's different
Trains a model and reports accuracy.
Detects known resistance genes from a sequence.
Grounds the model + feature choices in evidence, prevents clonal/sample-id leakage with group-aware CV, ships a model card, and enforces the research-only boundary in the type system.
How a model is built
- 1 · Bring the data
Pick an upstream execution (the closed-loop lineage and its literature consensus thread through) and supply the feature matrix + outcome, or load an illustrative example.
- 2 · Train, leakage-aware
Omnix trains the chosen model family with group-aware cross-validation, so isolates from the same clonal lineage (or repeated samples) never straddle train and test. Every reported metric is out-of-fold.
- 3 · Explain and ground
SHAP gives global and per-prediction attribution; for AMR, top features are checked against the resistance-determinant catalogues; model and feature choices are tied to the literature consensus where available.
- 4 · Card and disclose
A model card, purpose, training data, performance, limitations, failure modes, exports to Markdown and PDF, with the research-only boundary stated and the human accountable.
The decision-support boundary
The research-only boundary is not a checkbox you can click past. A model card cannot be constructed without an interpretability attribution and citation evidence for both the model and the features, and its intended-use value can only be research-only, enforced at the data-validation layer, before anything is saved. Omnix is a research instrument: it surfaces predictions and the evidence behind them; it never makes or auto-executes a clinical, policy, or institutional decision.
What you get
- A trained model with out-of-fold cross-validation and a held-out test, against a baseline.
- Mandatory SHAP feature attribution, global and per-prediction, with AMR-determinant grounding.
- Probability calibration (reliability curve, Brier, ECE) where applicable.
- Proactive AI flags: overfitting, class imbalance, small-cohort external-validation prompts.
- A model card exportable to Markdown and PDF, plus an electronic signature and full provenance.