Home Heroes De Novo Discovery Gene Scan CDx Pathway ♻️ Repurposing 🚦 Gates Models Quantum 🧪 Wet Lab Schedules 🎛️ Control 🧬 CIFOG 📈 Milestones 🧬 BioBrain 🧠 NPEAD Architecture 🔬 RODE 🧠 Cortex-R Workspace 🧠 V2 🧠 V3 📦 Knowledge 💰 Revenue 🧪 Tests 📊 Costs 📚 Datasets

📚 Data Sources & API Access

All genomic data sources — access instructions, API keys, and registration links

Tier 1 — Active & Ready (Programmatic API)

Source	Genes	Access	API Endpoint	Status
TCGA/GDC The Cancer Genome Atlas	~20,000	Free — open access	`https://api.gdc.cancer.gov`	ACTIVE
DepMap/CRISPR Broad Institute	~18,000	Free — open access	`https://depmap.org/portal/download/all`	ACTIVE
Open Targets EBI + Wellcome Sanger	60,000+	Free — GraphQL API	`https://api.platform.opentargets.org/api/v4/graphql`	ACTIVE
Human Protein Atlas KTH / Uppsala	20,000	Free — REST API + bulk download	`https://www.proteinatlas.org/api`	READY
GTEx NIH Common Fund	~56,000	Free — REST API	`https://gtexportal.org/api/v2`	READY
ChEMBL EMBL-EBI	15,000+	Free — REST API	`https://www.ebi.ac.uk/chembl/api/data`	ACTIVE
ClinicalTrials.gov NIH/NLM	All	Free — REST API v2	`https://clinicaltrials.gov/api/v2/studies`	ACTIVE
STRING PPI EMBL / SIB / CPR	~20,000	Free — REST API	`https://string-db.org/api`	ACTIVE
KEGG Pathways Kyoto University	~8,000	Free — REST API	`https://rest.kegg.jp`	ACTIVE
PubChem NCBI	110M+ compounds	Free — PUG REST	`https://pubchem.ncbi.nlm.nih.gov/rest/pug`	ACTIVE
UniProt EBI / SIB / PIR	250M+ sequences	Free — REST API	`https://rest.uniprot.org`	ACTIVE

FGAO Fragment Memory Bank

Source	Fragments	Authority	Use	Status
ChEMBL Approved Drugs max_phase=4, drug-sized	215	0.65–1.0 (boosted)	Fragment recombination for de novo design	ACTIVE
Kinase Inhibitor Hits CDK2/BRAF/ALK actives, pChEMBL≥7	30+	0.5–0.9	Target-specific scaffold generation	ACTIVE
Docking Outcomes (self-learned) GPU pipeline feedback loop	112	0.0–1.0	Validated fragments from successful dockings	ACTIVE

Total fragment bank: 233 fragments · DynamoDB: fgao-fragment-memory · Updated nightly via Phase 12/12b

CortexQ Generative Model (De Novo)

Version	Architecture	Training Data	Validity	Features	Status
v4 (current) Jun 20, 2026	3.2M param Transformer 6 layers · 8 heads · 256 dim	10,258 molecules	100%	SELFIES + Pathway + Pocket + Target Class conditioning	ACTIVE
v3	3.2M param GPT	10,258 (RL-weighted)	82%	SMILES + winners fine-tuned + docking reward	ARCHIVE

v4 Training Dataset Composition

Source	Molecules	Selection Criteria	Weight
ChEMBL Drug-Like Kinase inhibitors, epigenetic modulators, approved drugs	2,677	pChEMBL≥6, MW 200–600, max_phase≥3	1.0×
Hero Docking Winners Pipeline-validated compounds (Vina ≤ -6.5)	2,449	GPU-validated, passed NPEAD gates	1.5× (reward-weighted)
FGAO Fragment Expansions High-authority fragments from memory bank	224	Authority ≥ 0.5, recombined	1.0×
RL-Augmented (Docking Reward) Upsampled by Vina score (better dock → more copies)	4,908	Top performers duplicated by reward signal	2.0–3.0× for Vina ≤ -8.0
Total: 10,258 training molecules

v4 Model Parameters

Parameter	Value	Parameter	Value
Parameters	3.2M	Format	SELFIES (100% valid)
Layers	6 transformer blocks	Vocab Size	135 tokens
Attention Heads	8	Context Length	128 tokens
Embedding Dim	256	Model Size	12.5 MB
Epochs Trained	40	Final Loss	0.59
Learning Rate	3e-4 (cosine decay)	Batch Size	64
Optimizer	AdamW (β1=0.9, β2=0.999)	Dropout	0.1

Conditioning Tokens (23 total)

Pathway (11):  <MAPK> <PI3K> <p53> <WNT> <TGFb> <EGFR> <JAK_STAT> <NFkB> <Hypoxia> <Androgen> <Estrogen>
Target (8):    <kinase> <epigenetic> <phosphatase> <protease> <gpcr> <nuclear> <transporter> <other>
Pocket (4):    <pocket_deep> <pocket_shallow> <pocket_allosteric> <pocket_unknown>

Rejection Sampling (Self-Improving Nightly Loop)

Step	Count	Description
1. Generate	1,000	Sample from current model with pathway+pocket+class tokens
2. Filter	~800	SELFIES validity (100%) + Lipinski + PAINS + 30 med-chem patterns
3. Dock	50	Top 50 by QED+SA → GPU AutoDock Vina
4. Select	10	Best docking scores (Vina ≤ -6.5 threshold)
5. Retrain	—	Fine-tune model on winners (reward-weighted)

Result: 5/7 novel pancreatic targets validated with generative molecules (MAPK4, SKA1, CXXC1, CFAP53, MEX3C) — zero prior art existed for these targets.

Knowledge Capsules (Brain Intelligence Layer)

Brain	Capsules	Cancer Types	Source	Status
Genomics DepMap essentiality + expression	5	Prostate, Pancreatic, Lung, Breast, TNBC	cortex-heroes + DepMap	ACTIVE
Chemistry Vina scores + known inhibitors	5	All cancer types	cortex-molecules + ChEMBL	ACTIVE
Transcriptomics Differential expression (LFC)	5	All cancer types	TCGA + GEO	ACTIVE
Clinical Survival + staging + trials	4	Prostate, Pancreatic, Lung, Breast	ClinicalTrials.gov	ACTIVE
Pathway KEGG + pathway mapping	4	Major cancer types	KEGG + OpenTargets	ACTIVE
Biomarkers CDx potential + companion diagnostics	4	Major cancer types	cortex-heroes CDx fields	ACTIVE
Safety Chemistry flags + selectivity	4	Major cancer types	30 med-chem patterns (Brenk/Lilly)	ACTIVE
Pharmacology QED drug-likeness scores	3	Major cancer types	RDKit QED + SA scoring	ACTIVE
Evidence Justification bundles	2	Prostate, Pancreatic	Brain-bus evidence chains	ACTIVE
Proteomics AlphaFold structures + UniProt	2	Prostate, Pancreatic	AlphaFold + UniProt	ACTIVE

Total: 38 capsules across 10 brain types × 5 cancer types · DynamoDB: cortex-brain-bus (pk: capsule:{brain}:{cancer}) · Generated nightly by Step 17d · Growing with each cycle

Capsule → CortexQ Generator Feedback (Live)

Capsule Signal	Generator Effect	Mechanism	Status
Chemistry (Best Vina ≤ -8.0)	Lower temperature → exploit mode	temp 0.85 → 0.75 (tighter generation around known good space)	ACTIVE
Safety (all clear)	Maintain current patterns	No constraint tightening needed — existing chemistry passes gates	ACTIVE
Evidence (weak justification)	Higher temperature + more candidates	temp +0.05, n ×1.2 — explore novel space, generate more to find hits	ACTIVE
Transcriptomics (many dysregulated)	Prioritize targets with genomic signal	HYPOTHESIS capsule_bias flag read by gpu-batch-rs routing	ACTIVE

Both generators (cortexq_generate.py v4 SELFIES + cortex_gen_v3.py heuristic) read capsules at runtime. Capsule intelligence is the compressed output of 15 brains feeding back into molecule generation.

Tier 2 — Curated Institutional

Source	Data	Access	Registration	Status
AACR Project GENIE 19 cancer centers, 100K+ tumors	Real-world genomics	Free — registration required	Synapse registration →	PENDING
cBioPortal MSK, DFCI, multi-center	Curated multi-omics	Free — REST API	`https://www.cbioportal.org/api`	READY
COSMIC Wellcome Sanger Institute	1.5M somatic mutations	Free academic / Paid commercial	Register for API key →	PENDING
GEO/NCBI National Library of Medicine	4M+ expression profiles	Free — E-utilities API	Get API key (optional) →	READY

Tier 3 — Institutional Agreement Required

Source	Data	Access	How to Apply
Genomics England	100K genomes (rare + cancer)	Institutional agreement	Research application →
UK Biobank	500K participants, longitudinal	Institutional agreement	Apply for access →
ICGC/PCAWG	Whole-genome pan-cancer	DACO application	DACO application →
CPTAC	Proteogenomics	Free — GDC portal	`https://pdc.cancer.gov/pdc/`
PharmGKB	Pharmacogenomics	Free — Creative Commons	`https://api.pharmgkb.org`

API Access Instructions

Open Targets (recommended next)

# GraphQL — no auth needed
curl -X POST https://api.platform.opentargets.org/api/v4/graphql \
  -H "Content-Type: application/json" \
  -d '{"query":"{target(ensemblId:\"ENSG00000141510\"){approvedSymbol tractability{label modality value}}}"}'

Human Protein Atlas

# REST — no auth needed
curl "https://www.proteinatlas.org/api/search_download.php?search=ADAM28&format=json&columns=g,t,up,rnats,rnac"
# Bulk: https://www.proteinatlas.org/download/proteinatlas.tsv.zip (~100MB)

COSMIC (requires API key)

# 1. Register: https://cancer.sanger.ac.uk/cosmic/register
# 2. Get API key from account settings
# 3. Store key: edice-platform-state pk=creds:cosmic
curl -H "Authorization: Bearer YOUR_KEY" "https://cancer.sanger.ac.uk/api/v1/census/genes"

AACR GENIE (requires Synapse account)

# 1. Create account: https://www.synapse.org/register
# 2. Accept data use agreement for syn7222066
# 3. Download via synapseclient:
pip install synapseclient
synapse get syn7222066 --recursive

NCBI GEO (optional API key for higher rate)

# Without key: 3 req/sec. With key: 10 req/sec
# Get key: https://www.ncbi.nlm.nih.gov/account/settings/#/api-key
curl "https://eutils.ncbi.nlm.nih.gov/entrez/eutils/esearch.fcgi?db=gds&term=prostate+cancer&retmode=json&api_key=YOUR_KEY"

Backup Location

s3://edicer-dev-research-objects-865950114883/data-sources/
├── opentargets_cancer_sample.json    (16,737 cancer targets)
├── proteinatlas.tsv.zip              (20K proteins, expression)
├── chembl_targets_sample.json        (protein targets)
├── clinicaltrials_oncology_sample.json (oncology trials)
├── cbioportal_studies.json           (100 studies)
└── opentargets_count.json            (total count)

MindNetQ Inc. — Computational Oncology & Drug Discovery Division