An interactive sandbox for exploring RNA‑guided DNA editing, mutation detection, and computational drug ranking — driven by real client‑side logic, not a live lab.
⚠ Illustrative model — synthetic patients, sequences & scores for teaching. Not a real trial, EHR, or blockchain network.
Real-time · Interactive · Backend-connected
Molecule Construction Lab
Place atoms, wire up bonds, and build a real molecule from scratch — validated with actual chemistry (RDKit via the Python backend, with a labeled local fallback), then screen it against whatever gene the disease analyzer currently flags. Design a drug candidate starting from the DNA, not the other way around.
01Image intake triage — YOLOv8 object detection
A real, pretrained YOLOv8n checkpoint runs live in your browser (TensorFlow.js) as a first-pass scan of an uploaded photo, before you move deeper into the sequence, disease, and drug panels below.
Upload a photo to run live object detection — nothing leaves your browser.
AI summary
Real YOLOv8n weights, converted to TensorFlow.js, detecting the 80 general-purpose COCO classes (people, animals, vehicles, everyday objects). It is not trained on microscopy, histology, or pathogen imagery — this is a genuine vision-triage stage feeding the pipeline below, not a diagnostic image classifier.
Docks the selected compound's own 3D structure (generated live from its SMILES) into a stylized pocket for either the disease gene currently flagged, or a real microbial drug target — then searches candidate orientations for the best geometric fit.
All 8 synthetic compounds plotted by their real RDKit descriptors — molecular weight, LogP, and polar surface area — so you can see how "drug-likeness" clusters in 3D rather than just as a ranked list.
drag to rotate · scroll to zoom
high likeness (QED‑lite ≥ 70)mid likeness (45–69)low likeness (< 45)
Axes: X = molecular weight (150–420), Y = LogP (−4 to 4), Z = TPSA (40–170 Ų) — real values from the RDKit descriptor table, min–max scaled to fit the plot. Point color is the same deterministic QED‑lite index used in the drug screening panel above.
Generates a short synthetic sequencing read for every enrolled patient (reference sequence plus a small simulated error/mutation rate), aligns each read to the reference position-by-position, and calls variants from real mismatches — the same base logic behind a toy read-alignment pipeline, run entirely in this tab.
referencematches referencecalled SNP
Reads are generated from patients in the clinical trial cohort below — enroll patients there first, or click "Generate reads" to auto-create a small default cohort. Each simulated read gets a ~4% per-base error rate; positions where 2 or more reads disagree with the reference are called as variants (a deliberately simplified stand-in for real variant callers like GATK or DeepVariant).
09CRISPR knowledge assistant — local retrieval (RAG‑lite)
Ask a question about CRISPR biology, mechanism, or applications. This runs a genuine TF‑IDF + cosine‑similarity search over a small built‑in knowledge base, then composes an answer by extracting the top‑matching sentences — a real retrieval step, honestly without a generative LLM behind it (a static file like this has no server or API key to call one).
1987Nakata et al., Osaka University — CRISPR repeats first reported in E. coli
~2005Spacer sequencing links CRISPR to a prokaryotic adaptive immune function
2012Doudna & Charpentier fuse crRNA + tracrRNA into a single programmable sgRNA
2020Nobel Prize in Chemistry awarded for CRISPR-Cas9 genome editing
How was CRISPR discovered?How does Cas9 recognize its target?HDR vs NHEJ?Clinical applications?
Retrieval: TF‑IDF vectors + cosine similarity over – short passages
10Clinical trial simulator
12Distributed audit ledger
Every edit, consent, and adverse event is hashed with real SHA‑256 and chained to the previous block — the same tamper‑evidence idea behind Hedera's Consensus Service. This runs locally in your browser; it is not a live call to the Hedera network.
11Clinical outcome simulator — survival analysis
Simulates event-free survival for a treatment arm and a control arm from synthetic exponential survival times (treatment hazard reduced in proportion to the top compound's fit score), then computes a real Kaplan‑Meier curve and a real log‑rank test statistic on that simulated data.
Treatment armControl arm
Log‑rank p‑value—
This is a real, if simplified, statistical pipeline (Kaplan‑Meier estimator + log‑rank chi‑square, exact for 1 degree of freedom via the χ²↔normal identity) applied to simulated exponential survival times, not real patient outcomes. Re-run it after changing the detected mutation or repair strategy above to see the response‑rate assumption shift.
What's real vs. simplified here
This demo runs entirely in your browser — no external AI model, genome database, EHR system, or live blockchain network is called.
3D viewer: real Three.js geometry, procedurally built from the sequence you edit.
gRNA / PAM search: real string matching against the sequence shown.
Disease matches: a small hand‑written reference table (6 entries), not ClinVar/OMIM/gnomAD.
Drug descriptors: real RDKit‑computed molecular weight, LogP, H‑bond donors/acceptors, TPSA and rotatable bonds for fictional molecules with valid SMILES — see the companion Python script. The 2D structures are drawn live from those same SMILES strings with SmilesDrawer, a genuine chemical structure parser.
Likeness score: a real, deterministic index computed from the descriptors above (QED‑style desirability curves for MW/LogP/HBD/HBA/TPSA/rotatable bonds) — not random, but still a simplified drug‑likeness heuristic, not a measured property.
Dock fit score: no longer a seeded placeholder — every compound in the screening list now runs through the exact same pipeline as the binding-pose simulator below (SMILES → 3D atom embedding → pocket generation → 18-orientation pose search), automatically, against the currently flagged gene. It's a real simulation output, cached per-compound so re-scanning stays fast — but it's still the same simplified, rule-based geometric scoring described below, not a validated binding free energy.
Binding pose simulator: the ligand's 3D shape is genuinely generated from its SMILES (parsed atom-by-atom, then relaxed into 3D with a real bonded-spring/repulsion algorithm) and the best of 18 candidate orientations is picked by a real geometric-contact scoring function. The receptor pocket is a stylized, deterministic point-cloud — not a real protein structure from the PDB — so treat the pose and score as an illustration of docking concepts (donor/acceptor/hydrophobic complementarity, pose search), not a substitute for physics-based docking (AutoDock, Glide) against a real crystal or cryo-EM structure.
Clinical trial cohort: entirely synthetic, randomly generated patients — no real people, no real PII, no real regulatory submission.
Audit ledger: a genuine local SHA‑256 hash chain (you can verify it below) illustrating what a distributed ledger secures — not a connection to Hedera testnet/mainnet, which needs an account, HBAR, and the Hedera SDK running server‑side. The companion Python script includes real, runnable example code for that integration.
Image intake triage: a genuine pretrained YOLOv8n checkpoint runs live in your browser via TensorFlow.js — real weights, real inference, nothing uploaded to a server. It detects the 80 general‑purpose COCO classes (people, vehicles, animals, household objects); it is not trained on microscopy, histology, or pathogen imagery, so it's a real demonstration of a vision‑triage stage, not a diagnostic image classifier.
3D chemical space: the same real RDKit descriptors (MW, LogP, TPSA) plotted as real 3D coordinates, min‑max scaled — a genuine data visualization, just of a small (8‑compound) synthetic set.
CRISPR knowledge assistant: real TF‑IDF vectorization + cosine similarity search over a small, original 15‑passage knowledge base, followed by extractive sentence selection — genuine retrieval and ranking math. There is no generative LLM behind it: a static HTML file has no server or API key to call one, so answers are assembled from retrieved sentences rather than written fresh.
13Final output dashboard
A live checklist of what this session has actually produced — updates as you use the panels above.
Molecule Construction Lab
Click the canvas to place an atom of the selected element. Turn on bond mode, pick a bond order, then click two atoms to connect them. Build & score sends the graph to the Python backend (real RDKit) — if it's unreachable, a clearly-labeled local heuristic estimate is shown instead so the lab still works standalone.
Place a few atoms, connect them, then click "Build & score" to run real chemistry validation.