CEREBRO-X
CEREBRO-X
AI-Native · Physics-Informed · Molecule-Aware
An open-source computational engine for CNS drug delivery design. 62 scoring criteria, every computed value carrying its method and citation. Built for researchers, pharma R&D, and academic collaborators — a research prototype, not a validated clinical tool.
† Pardridge WM (2005) NeuroRx 2:3 — a literature figure, not a CEREBRO-X result.
The molecule is rarely the problem. The delivery system is. Researchers spend months designing nanocarriers that collapse, get protein-coated, or simply cannot cross the BBB. Each failure costs hundreds of thousands.
What if you could simulate every failure mode before synthesis?
Not a chatbot. Not an LLM wrapper. CEREBRO-X is the reverse of legacy platforms: remove the AI and the engine stops.
57 ML models predict drug-DDS properties in milliseconds. Heavy physics is invoked only on uncertainty.
A GA optimizer proposes new DDS formulation candidates and scores each generation against the same 62-criterion rubric.
A weighted composite ranks all candidate DDS by predicted performance, with per-criterion breakdown.
A Keras model trained on the public BBBP dataset (RDKit Morgan fingerprints) classifies BBB permeability for small molecules.
Missing property? Resolved through a tiered cascade — public databases, cheminformatics, then physics fallback — every value tagged with its tier and confidence.
FastAPI + Docker Compose. Runs locally today; designed to scale to a cloud deployment.
Six independent computational layers, each with a single responsibility. Replaceable, scalable, isolated. A mesh of specialists — not a monolith.
SMILES/FASTA validation, canonicalisation, multi-drug parsing, property tier-resolution.
57 XGBoost / GNN / QSAR models — each predicts properties + uncertainty for one principle.
AutoDock Vina docking, a 3-compartment PBPK ODE integrator (scipy), protein-corona and glymphatic-clearance models.
DAG executor coordinating principle execution, weight aggregation, group rollups, composite scoring.
Interactive HTML5, publication-ready PDFs, MP4 videos, Excel with provenance, FDA Pre-IND drafts.
FastAPI gateway, Celery queue, Redis cache, Postgres metadata, active-learning loop, Prometheus monitoring.
Hover any principle to read its method, references, and class. Three classes — one unified score.
From quantum tunnelling to glymphatic clearance, protein corona to FDA Pre-IND. Eight CNS pharma dimensions, 62 mathematical models, one composite verdict.
A CNS-specific PBPK model simulates the drug's full journey. Output: time-resolved concentrations in every compartment, every hour.
IV/SC/oral kinetics
Tight-junction crossing
Target engagement
CSF clearance
Hepatic + renal
Two Alzheimer's drugs, same indication. CEREBRO-X identified different optimal carriers from molecular properties alone — real output from today's pipeline run, verdicts included as-is.
| Property | Donepezil | Rivastigmine |
|---|---|---|
| Molecular Weight | 379.5 Da | 250.3 Da |
| LogP | 4.77 (lipophilic) | 2.30 (hydrophilic) |
| Top-1 DDS | RVG29-PLGA-NP | RVG29-PLGA-NP |
| Composite (Class A) | 80.7 / 100 EXCELLENT | 83.2 / 100 EXCELLENT |
| Class B Deep Validation | PASSED 20/28 | PASSED 20/28 |
No LLMs in the scoring path. Every score comes from a mathematical model, cheminformatics computation, or ML regressor — each with a cited method.
A 7-tier resolver cascade — researcher override, public databases, cheminformatics, physics fallback — computes a value and tags it with tier + confidence.
Free, open-source, MIT licensed. Clone the repo, run it locally, or open an issue/PR — no request process required.
Run it entirely on your own infrastructure — the engine never phones home. Bring your own data locally; nothing is uploaded anywhere by default.
Applied to 6 real drugs (3 molecule classes). Class A surrogate scores are internal heuristics; Class B deep-physics re-validation disagreed with 3 of them — reported, not hidden.
Runtime scales with candidate-DDS count and whether deep-physics fallback triggers; the surrogate-only pass is the fast path.
RDKit Morgan/ECFP4 fingerprints, Henderson-Hasselbalch ionisation profiles, and a Keras DNN trained on the public BBBP dataset.
Drafts a Pre-IND-style framework and an FTO/patentability score for the Top-1 candidate — a starting outline, not a regulatory submission.
Today. 62-criterion pipeline with Class A + Class B validation. Local Docker deploy. Excel → reports.
Next. Web UI for researchers. SMILES in, dashboards out. No Excel needed.
Longer-term idea, not built. Multi-tenant cloud.
Trained on public datasets only — no proprietary or pharma-scale compound libraries were used. If your data could improve it, the code is open: fork it, retrain on your own library, and keep the results.
Free, MIT licensed, no conditions attached.
git clone the repo, install requirements.txt, run it on your own machine.
Read the engineering audit for known limitations before you rely on any output.
Fill the Excel template with SMILES/FASTA + candidate DDS, run the pipeline locally.
File issues, open pull requests, or fork it for your own research.
Free · MIT licensed · no fees, no conditions, no required co-authorship
Built by Muhammad Talaat · BPharm, R&D Computational Lead
Ain Shams University · Faculty of Pharmacy
mohamed.talaat@pharma.asu.edu.eg
© 2026 MUHAMMAD TALAAT · MIT LICENSED · OPEN SOURCE