Update README: 1.02B compounds (added 99M CPU screen results)

README.md

@@ -31,14 +31,34 @@ dataset_info:
       dtype: float64
     - name: cns_mpo_score
       dtype: float64
+    - name: efflux_MDR1_HUMAN
+      dtype: float64
+    - name: efflux_A4D1D2_HUMAN
+      dtype: float64
+    - name: efflux_ABCG2_HUMAN
+      dtype: float64
+    - name: efflux_MRP1_HUMAN
+      dtype: float64
+    - name: efflux_MRP2_HUMAN
+      dtype: float64
+    - name: efflux_MRP4_HUMAN
+      dtype: float64
+    - name: efflux_S47A1_HUMAN
+      dtype: float64
+    - name: efflux_S22A8_HUMAN
+      dtype: float64
     - name: p_bbb
       dtype: float64
+    - name: heuristic_reason
+      dtype: string
+    - name: heuristic_veto
+      dtype: string
     - name: score_total
       dtype: float64
   splits:
     - name: train
-      num_examples: 924220136
-  dataset_size: 126000000000
+      num_examples: 1023158502
+  dataset_size: 140000000000
 license: cc-by-nc-4.0
 task_categories:
   - tabular-classification
@@ -50,25 +70,36 @@ tags:
   - cns
   - molecular-properties
   - screening
-pretty_name: "BBB-Nuke: 924M Compound BBB Permeability Screen"
+pretty_name: "BBB-Nuke: 1B Compound BBB Permeability Screen"
 size_categories:
   - 1B<n<10B
 ---
 
-# BBB-Nuke: 924M Compound Blood-Brain Barrier Permeability Screen
+# BBB-Nuke: 1 Billion Compound Blood-Brain Barrier Permeability Screen
 
 ## Overview
 
-924 million small molecules screened for blood-brain barrier (BBB) permeability using the [BBB-Nuke](https://github.com/ATTN-Lab) pipeline.
+1.02 billion small molecules screened for blood-brain barrier (BBB) permeability using the [BBB-Nuke](https://github.com/ATTN-Lab) pipeline (v0.12.0).
+
+## Screening Pipeline
+
+Each compound was scored through:
+1. **Standardization** - SMILES canonicalization via RDKit
+2. **Physicochemical properties** - MW, LogP, TPSA, HBD, HBA, Fsp3, heavy atom count
+3. **pKa prediction** - Representative pKa via MolGpKa (batched GCN inference)
+4. **CNS-MPO scoring** - Multi-parameter optimization (6 properties)
+5. **Efflux transporter prediction** - 8 efflux proteins (S1 fingerprint RF classifiers)
+6. **Classifier + heuristic** - Final P_BBB probability score (5-fold CV AUROC 0.933)
 
 ## Data Sources
 
-| Source | Files | Compounds |
-|--------|-------|-----------|
-| Enamine REAL | 2,000 | ~199M |
-| Existing (ZINC/ChEMBL) | 7,384 | ~666M |
-| PubChem | 625 | ~59M |
-| **Total** | **10,009** | **924,220,136** |
+| Source | Parquet Dir | Compounds |
+|--------|-------------|-----------|
+| Enamine REAL |  | ~199M |
+| Existing (ZINC/ChEMBL) |  | ~666M |
+| PubChem |  | ~59M |
+| CPU Screen (ZINC) |  | ~99M |
+| **Total** | | **1,023,158,502** |
 
 ## Key Columns
 
@@ -77,11 +108,14 @@ size_categories:
 -  - CNS multi-parameter optimization score
 - , ,  - Key physicochemical descriptors
 -  - Predicted pKa
+-  - Efflux transporter affinity scores (8 proteins)
 -  - Data source (enamine_real, existing, pubchem)
+-  - Composite heuristic score
 
 ## Compute
 
-Screened on Azure ML using 8x NVIDIA A100 80GB GPUs (Standard_ND96amsr_A100_v4).
+- **924M compounds**: Azure ML, 8x NVIDIA A100 80GB GPUs (Standard_ND96amsr_A100_v4)
+- **99M compounds**: Azure ML, 64-core CPU (Standard_E64ds_v4), batched pKa + S1 efflux only (no PSICHIC)
 
 ## Citation