PROJECT IRIS

Cloning Feasibility Model (CFM)

Probabilistic gate-based model of SCNT (Somatic Cell Nuclear Transfer) that formalizes end-to-end cloning feasibility as a composition of biological stage constraints with uncertainty quantification.

• Gate structure: SCNT decomposed into stage-wise gates (e.g., reprogramming, cleavage, implantation, placental development) to expose where variance accumulates.
• Sensitivity layer: Computes gate–outcome associations across draws to identify the strongest bottlenecks (targets for intervention and model refinement).
• Outputs: Full outcome distribution + uncertainty bands (P05/P50/P95) suitable for reporting and regression tests across versions.

BSE Interactive PK Demo

Three-compartment mechanistic PK model solved via RK4 integration, parameterized from literature (Marsh model). Interactive dosing + infusion controls with live concentration curves and metrics.

Bio-Simulation Engine (BSE)

Mechanistic “Virtual Physiology” engine modeling organs as coupled dynamical systems (compartments with explicit state variables) rather than black-box predictors.

• Architecture: Multi-compartment PK/PD with explicit state evolution per organ (concentration, binding fraction, immune activity, toxicity thresholds).
• Coupling: Inter-organ flows model distribution and exchange (circulation-driven transport + organ-specific transformation/clearance).
• ADME backbone: Absorption, distribution, metabolism, excretion encoded as parameterized transfer/clearance operators across compartments.
• Modularity: Organ plug-ins enable incremental scope growth (add organs, immune overlays, toxicity models) without rewriting the core solver.
• Current experiment: Rabbit → Human scaling tests using physiology/allometric transformations of organ volumes, flow rates, and clearance constants to evaluate cross-species generalizability.

Neural Self-Consistent System (NSCS)

Closed-loop model discovery framework that ingests time-series trajectories, proposes symbolic candidates, rejects unstable forms, and applies neural residual correction to improve stability and predictive fidelity.

• 1) Ingest: read trajectories (states over time), normalize channels, define target variables and candidate feature library.
• 2) Symbolic candidate generation: generate ODE candidates via constrained symbolic regression / structured search (candidate families, parameter sweeps).
• 3) Score + reject: evaluate candidates on fit + stability (e.g., exploding/oscillatory divergence, nonphysical trends) and discard unstable solutions.
• 4) Residual correction: train a neural residual model Δ(t, x) to learn the mismatch between mechanistic candidate predictions and observed trajectories.
• 5) Compose: final model = mechanistic form + residual correction (only if it improves stability/generalization) with versioned metrics.
• 6) Feedback: inject improved equation/parameters back into BSE to iteratively refine the simulator under the same evaluation protocol.

Fast Radio Burst Scoring & Evaluation

A reproducibility-first pipeline that ingests FRB catalogs and derived features, produces event scores, and logs evaluation metrics per run.

• Ingest: standardized catalog tables with version tags.
• Feature layer: derived burst features and normalization.
• Model layer: per-event scoring with versioned outputs.
• Evaluation: AUC, PR-AUC, F1, and threshold sweeps.
• Run ledger: each experiment recorded as a row in a CSV.

Downloadable Artifacts

Versioned outputs from experiments. Update the run ledger first, then upload matching figures and optional ZIP bundles.

• Run ledger fields: run_id, date, model, seed, AUC, PRAUC, F1, notes.
• Portal pack: latest ledger + figures + short summary README.