Spaces:
Running
Running
File size: 10,109 Bytes
3cc173b | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 | """
Applicability Domain (AD) Checker using One-Class SVM
Uses One-Class SVM on GNN embeddings to determine if a molecule
is within the training distribution (applicability domain).
"""
import numpy as np
from sklearn.svm import OneClassSVM
from sklearn.preprocessing import StandardScaler
from typing import List, Tuple, Optional
import pickle
from pathlib import Path
class ApplicabilityDomainChecker:
"""
Check if molecules are within the applicability domain using One-Class SVM.
The idea:
1. Train on embeddings from training set
2. For new molecules, check if they're in the same latent space region
3. Molecules outside AD → predictions are unreliable
"""
def __init__(self, nu: float = 0.02, kernel: str = 'rbf', gamma: str = 'scale'):
"""
Args:
nu: Fraction of training data to be treated as outliers (0.01-0.05)
Lower nu = stricter AD, higher nu = more permissive
Typical: 0.02 (2% outliers)
kernel: 'rbf' (default), 'linear', 'poly', 'sigmoid'
gamma: Kernel coefficient. 'scale' (1/(n_features*X.var())) recommended
over 'auto' (1/n_features) for high-dimensional embeddings
"""
self.nu = nu
self.kernel = kernel
self.gamma = gamma
self.scaler = StandardScaler()
self.oc_svm = OneClassSVM(
nu=nu,
kernel=kernel,
gamma=gamma
)
self.is_fitted = False
# Score range from training set — used to normalise confidence to 0-100
self.score_min = None
self.score_max = None
def fit(self, train_embeddings: np.ndarray):
"""
Fit the One-Class SVM on training set embeddings.
Args:
train_embeddings: Array of shape (n_train, embedding_dim)
"""
print(f"Training Applicability Domain checker...")
print(f" Training samples: {len(train_embeddings)}")
print(f" Embedding dimension: {train_embeddings.shape[1]}")
print(f" Nu (outlier fraction): {self.nu}")
# Normalize embeddings
embeddings_scaled = self.scaler.fit_transform(train_embeddings)
# Fit One-Class SVM
self.oc_svm.fit(embeddings_scaled)
# Store training score range for normalised confidence
train_scores = self.oc_svm.decision_function(embeddings_scaled)
self.score_min = float(train_scores.min())
self.score_max = float(train_scores.max())
# Check how many training samples are classified as outliers
train_predictions = self.oc_svm.predict(embeddings_scaled)
n_outliers = (train_predictions == -1).sum()
print(f" Training outliers: {n_outliers}/{len(train_embeddings)} ({n_outliers/len(train_embeddings)*100:.1f}%)")
print(f" Decision score range: [{self.score_min:.4f}, {self.score_max:.4f}]")
self.is_fitted = True
print("✓ AD checker ready!")
def predict(self, embeddings: np.ndarray) -> np.ndarray:
"""
Check if molecules are within the applicability domain.
Args:
embeddings: Array of shape (n_molecules, embedding_dim)
Returns:
predictions: Array of +1 (in AD) or -1 (out of AD)
"""
if not self.is_fitted:
raise RuntimeError("AD checker not fitted. Call .fit() first.")
# Normalize
embeddings_scaled = self.scaler.transform(embeddings)
# Predict
return self.oc_svm.predict(embeddings_scaled)
def decision_function(self, embeddings: np.ndarray) -> np.ndarray:
"""
Get decision scores (distance from AD boundary).
Higher score = more confident the molecule is in AD
Lower/negative score = outside AD
Args:
embeddings: Array of shape (n_molecules, embedding_dim)
Returns:
scores: Array of decision scores
"""
if not self.is_fitted:
raise RuntimeError("AD checker not fitted. Call .fit() first.")
embeddings_scaled = self.scaler.transform(embeddings)
return self.oc_svm.decision_function(embeddings_scaled)
def is_in_domain(self, embeddings: np.ndarray) -> np.ndarray:
"""
Boolean array: True if in AD, False if out of AD.
Args:
embeddings: Array of shape (n_molecules, embedding_dim)
Returns:
in_domain: Boolean array
"""
predictions = self.predict(embeddings)
return predictions == 1
def get_confidence_scores(self, embeddings: np.ndarray) -> np.ndarray:
"""
Convert decision scores to confidence scores (0-100).
Scores are normalised to the training set range so that:
- The most central training point scores 100
- The decision boundary scores 0
- Out-of-domain points score below 0 (clipped to 0)
Args:
embeddings: Array of shape (n_molecules, embedding_dim)
Returns:
confidence: Array of confidence scores (0-100)
"""
decision_scores = self.decision_function(embeddings)
# Normalise relative to the training score range recorded during fit()
score_range = self.score_max - self.score_min
if score_range < 1e-8:
# Degenerate case: all training scores identical
return np.where(decision_scores >= 0, 100.0, 0.0).astype(float)
confidence = (decision_scores - self.score_min) / score_range * 100
return np.clip(confidence, 0, 100)
def save(self, filepath: str):
"""Save the fitted AD checker."""
if not self.is_fitted:
raise RuntimeError("Cannot save unfitted AD checker")
with open(filepath, 'wb') as f:
pickle.dump({
'scaler': self.scaler,
'oc_svm': self.oc_svm,
'nu': self.nu,
'kernel': self.kernel,
'gamma': self.gamma,
'score_min': self.score_min,
'score_max': self.score_max,
}, f)
print(f"✓ AD checker saved to {filepath}")
@classmethod
def load(cls, filepath: str):
"""Load a fitted AD checker."""
with open(filepath, 'rb') as f:
data = pickle.load(f)
checker = cls(nu=data['nu'], kernel=data['kernel'], gamma=data['gamma'])
checker.scaler = data['scaler']
checker.oc_svm = data['oc_svm']
checker.score_min = data.get('score_min')
checker.score_max = data.get('score_max')
checker.is_fitted = True
print(f"✓ AD checker loaded from {filepath}")
return checker
# =============================================================================
# COMPLETE WORKFLOW
# =============================================================================
"""
STEP 1: Extract embeddings from training set
--------------------------------------------
from embedding_extractor import EmbeddingExtractor
from mixture_dcn_model import load_trained_model
# Load model
model = load_trained_model('models/mixture_dcn.pth')
# Create extractor
extractor = EmbeddingExtractor(model)
# Extract embeddings for ENTIRE training set
train_embeddings = extractor.extract_embeddings_from_smiles(
train_smiles,
featurizer=your_featurizer
)
STEP 2: Train One-Class SVM
---------------------------
from applicability_domain import ApplicabilityDomainChecker
# Create and train AD checker
ad_checker = ApplicabilityDomainChecker(
nu=0.02, # 2% outliers
kernel='rbf'
)
ad_checker.fit(train_embeddings)
# Save for later use
ad_checker.save('models/ad_checker.pkl')
STEP 3: Check new molecules during evolution
--------------------------------------------
# During evolution, for each batch of new molecules:
new_embeddings = extractor.extract_embeddings_from_smiles(
new_smiles_list,
featurizer=your_featurizer
)
# Check if in domain
in_domain = ad_checker.is_in_domain(new_embeddings)
# Get confidence scores
confidence = ad_checker.get_confidence_scores(new_embeddings)
# Filter molecules
for i, smiles in enumerate(new_smiles_list):
if in_domain[i]:
print(f"{smiles}: In AD (confidence: {confidence[i]:.1f}%)")
else:
print(f"{smiles}: OUT OF AD! Prediction unreliable!")
# Skip this molecule or flag it
STEP 4: Integrate into MixtureAwareMolecularEvolution
-----------------------------------------------------
See mixture_evolution_with_ad.py for full integration!
"""
# =============================================================================
# TUNING GUIDE
# =============================================================================
"""
How to choose 'nu' parameter:
-----------------------------
nu = 0.01 (1%) → Very strict AD, few molecules accepted
nu = 0.02 (2%) → Recommended starting point
nu = 0.03 (3%) → More permissive, good for exploration
nu = 0.05 (5%) → Very permissive
Start with 0.02 and adjust based on:
- If too many molecules are rejected → increase nu
- If you want stricter filtering → decrease nu
How to interpret decision scores:
---------------------------------
decision_function output:
> 0.5 : Strongly in AD (high confidence)
> 0.0 : In AD (medium confidence)
< 0.0 : Outside AD (low confidence)
< -0.5 : Strongly outside AD (very low confidence)
Confidence scores (0-100):
> 75 : High confidence - use prediction
50-75: Medium confidence - use with caution
25-50: Low confidence - unreliable
< 25 : Very low confidence - reject
""" |