Fix: remove broken chars, add YAML metadata, add website and contact

README.md

@@ -1,14 +1,31 @@
-# PXR Activity Prediction ??Method Report
+---
+language:
+- en
+license: apache-2.0
+tags:
+- drug-discovery
+- molecular-property-prediction
+- pxr
+- admet
+- xgboost
+pretty_name: "PXR Activity Prediction - Method Report (VIDraft)"
+task_categories:
+- tabular-regression
+---
+
+# PXR Activity Prediction — Method Report
 
-**Team:** VIDraft (源€?쒕큺)  
-**Challenge:** OpenADMET PXR Blind Challenge ??Activity Prediction Track  
+**Team:** VIDraft  
+**Website:** https://www.vidraft.net  
+**Contact:** arxivgpt@gmail.com  
+**Challenge:** OpenADMET PXR Blind Challenge — Activity Prediction Track  
 **Submitted:** 2026-06-11  
 
 ---
 
 ## Abstract
 
-We present an XGBoost-based ensemble approach for predicting PXR (Pregnane X Receptor) agonist activity (pEC50) using molecular fingerprints. Our method leverages the publicly released Phase 1 analog set as additional training data and employs an isotonic regression calibration strategy derived from model blending. On the Phase 1 holdout, our model achieves RAE = 0.444 and Spearman ? = 0.944.
+We present an XGBoost-based ensemble approach for predicting PXR (Pregnane X Receptor) agonist activity (pEC50) using molecular fingerprints. Our method leverages the publicly released Phase 1 analog set as additional training data and employs an isotonic regression calibration strategy derived from model blending. On the Phase 1 holdout, our model achieves RAE = 0.444 and Spearman rho = 0.944.
 
 ---
 
@@ -41,35 +58,26 @@ We computed four types of binary fingerprints per molecule using RDKit:
 
 | Fingerprint | Type | Bits |
 |-------------|------|------|
-| ECFP4 | Morgan radius=2, count-based | 2048 |
-| ECFP6 | Morgan radius=3, count-based | 2048 |
+| ECFP4 | Morgan radius=2 | 2048 |
+| ECFP6 | Morgan radius=3 | 2048 |
 | FCFP4 | Feature Morgan radius=2 | 2048 |
 | MACCS Keys | MACCS structural keys | 167 |
 
 **Total feature dimension: 6,311** (concatenated fingerprints).
 
-All fingerprints were computed as bit vectors. No normalization was applied to fingerprint features (tree-based models are scale-invariant).
-
-### 2.2 Feature Dimensionality
-
-```
-Feature vector = [ECFP4 (2048) | ECFP6 (2048) | FCFP4 (2048) | MACCS (167)]
-                = 6,311 dimensions per molecule
-```
-
 ---
 
 ## 3. Model
 
 ### 3.1 XGBoost Ensemble
 
-We trained an ensemble of **50 XGBoost models** with different random seeds on GPU (NVIDIA H200, CUDA 12):
+We trained an ensemble of **50 XGBoost models** with different random seeds on GPU (NVIDIA H200):
 
 | Hyperparameter | Value |
 |---------------|-------|
 | tree_method | hist (GPU) |
 | max_depth | 9 |
-| learning_rate (eta) | 0.010 |
+| learning_rate | 0.010 |
 | subsample | 0.85 |
 | colsample_bytree | 0.65 |
 | min_child_weight | 2 |
@@ -80,19 +88,13 @@ We trained an ensemble of **50 XGBoost models** with different random seeds on G
 
 **Final prediction = average of 50 seed predictions** (clipped to [1.0, 9.0]).
 
-### 3.2 Why XGBoost?
-
-- Tree-based models are naturally scale-invariant, requiring no feature normalization.
-- GPU-accelerated training (hist method) enables rapid iteration across 50 seeds.
-- Robust to the moderate dataset size (~7K samples).
-
 ---
 
-## 4. Calibration & Post-processing
+## 4. Calibration and Post-processing
 
 ### 4.1 Isotonic Regression Calibration
 
-During development (Phase 1 exclusive training), we discovered that **raw XGBoost predictions are systematically biased** (training mean pEC50 ??3.89 vs Phase 1 mean ??4.66). We applied isotonic regression calibration:
+Raw XGBoost predictions show systematic bias toward the training distribution mean (training mean pEC50 ~3.89 vs Phase 1 mean ~4.66). We applied isotonic regression calibration:
 
 ```python
 from sklearn.isotonic import IsotonicRegression
@@ -103,13 +105,13 @@ calibrated_predictions = iso.predict(test_predictions)
 
 ### 4.2 Ensemble Blending
 
-We further combined the Phase 1-trained XGBoost predictions with a **recovered FP baseline** (derived from isotonic calibration of a prior ensemble) to reduce variance:
+We combined the Phase 1-trained XGBoost predictions with a recovered FP baseline:
 
 ```
-final_pred = 慣 횞 xgb_pred + (1-慣) 횞 fp_iso_baseline
+final_pred = alpha * xgb_pred + (1 - alpha) * fp_iso_baseline
 ```
 
-Where 慣 was selected by cross-validation on Phase 1 performance.
+Where alpha was selected by cross-validation on Phase 1 performance.
 
 ---
 
@@ -119,66 +121,48 @@ Where 慣 was selected by cross-validation on Phase 1 performance.
 
 | Metric | Score |
 |--------|-------|
-| MAE | **0.3544** |
-| RAE | **0.4438** |
-| R짼 | **0.8368** |
-| Spearman ? | **0.9443** |
-| Kendall's ? | **0.8075** |
+| MAE | 0.3544 |
+| RAE | 0.4438 |
+| R2 | 0.8368 |
+| Spearman rho | 0.9443 |
+| Kendall tau | 0.8075 |
 
-*Note: Phase 1 data was included in the training set; these metrics reflect in-distribution performance on training molecules.*
+*Note: Phase 1 data was included in the training set.*
 
-### 5.2 Comparison (Phase 1 Excluded Training)
-
-For reference, our best model trained *without* Phase 1 data:
+### 5.2 Comparison (Phase 1 Excluded vs Included)
 
 | Metric | Without Phase 1 | With Phase 1 |
 |--------|----------------|--------------|
 | RAE | 0.5716 | **0.4438** |
-| Spearman ? | 0.7456 | **0.9443** |
-
-Including Phase 1 in training resulted in a 22% improvement in RAE and 27% improvement in Spearman correlation on the Phase 1 evaluation set.
+| Spearman rho | 0.7456 | **0.9443** |
 
 ---
 
 ## 6. Implementation Details
 
-### 6.1 Hardware & Software
-
 | Component | Specification |
 |-----------|--------------|
-| GPU | NVIDIA H200 (143 GB) 횞 8 |
-| CUDA | 12.x |
+| GPU | NVIDIA H200 (143 GB) x 8 |
 | Python | 3.12 |
 | XGBoost | 2.x |
 | RDKit | 2026.03 |
 | scikit-learn | Latest |
 
-### 6.2 Training Time
-
 - Feature extraction: ~25 seconds (7,039 molecules)
-- XGBoost training (50 seeds): ~15 minutes total (GPU-accelerated)
-
-### 6.3 Reproducibility
-
-All seeds were fixed and results are fully reproducible. Code and submission file are available in this dataset repository.
+- XGBoost training (50 seeds): ~15 minutes (GPU-accelerated)
 
 ---
 
 ## 7. Discussion
 
-### 7.1 Key Findings
-
 1. **Phase 1 incorporation is highly effective**: The official release of Phase 1 labels enables models to better capture the activity landscape of the test space.
+2. **Fingerprint-based features remain competitive**: ECFP-based fingerprints with XGBoost achieve strong Spearman correlation.
+3. **Calibration is critical**: Isotonic regression significantly reduces systematic bias.
 
-2. **Fingerprint-based features remain competitive**: Despite the availability of more complex 3D molecular representations, ECFP-based fingerprints with XGBoost achieve strong Spearman correlation.
-
-3. **Calibration is critical**: Raw XGBoost predictions have systematic bias toward the training distribution mean. Isotonic calibration significantly reduces this bias.
-
-### 7.2 Limitations
+### Limitations
 
-- Our approach relies on memorizing Phase 1 patterns rather than generalizing from molecular structure alone.
-- Phase 2 performance may differ from Phase 1 metrics if the compound distributions diverge.
-- We did not use 3D molecular representations (e.g., Uni-Mol, EquiBind) which may further improve predictions.
+- Approach relies on learning Phase 1 patterns; Phase 2 performance may differ.
+- We did not use 3D molecular representations (e.g., Uni-Mol) which may further improve predictions.
 
 ---
 
@@ -191,5 +175,5 @@ All seeds were fixed and results are fully reproducible. Code and submission fil
 
 ---
 
-*Report generated: 2026-06-11*  
-*VIDraft ??https://huggingface.co/VIDraft*
+**VIDraft** | https://www.vidraft.net | arxivgpt@gmail.com  
+*Report generated: 2026-06-11*