progress

tiger.py

@@ -99,17 +99,29 @@ def process_data(transcript_seq: str):
     return target_seq, guide_seq, model_inputs
 
 
-def prediction_transform(predictions: np.array, cutoff_path: str = 'cutoff.npy'):
-    """
-    :param predictions: in [0,1] where 0 represents most active guides
-    :param cutoff_path: full path to cutoff.npy (a float in [0,1] above which guides are inactive)
-    :return: predictions in [0,1] where 1 represents most active guides
-    """
-    cutoff = np.load(cutoff_path)
-    predictions[predictions > cutoff] = cutoff + (predictions[predictions > cutoff] - cutoff) * 0.01
-    predictions = predictions.max() - predictions
-    predictions = predictions / predictions.max()
-    return predictions
+def prediction_transform(predictions: np.array, **params):
+
+    # regime indices
+    active_saturation = predictions < params['a']
+    linear_regime = (params['a'] <= predictions) & (predictions <= params['c'])
+    inactive_saturation = params['c'] < predictions
+
+    # linear regime
+    slope = (params['d'] - params['b']) / (params['c'] - params['a'])
+    intercept = -params['a'] * slope + params['b']
+    predictions[linear_regime] = slope * predictions[linear_regime] + intercept
+
+    # active saturation regime
+    alpha = slope / params['b']
+    beta = alpha * params['a'] - np.log(params['b'])
+    predictions[active_saturation] = np.exp(alpha * predictions[active_saturation] - beta)
+
+    # inactive saturation regime
+    alpha = slope / (1 - params['d'])
+    beta = -alpha * params['c'] - np.log(1 - params['d'])
+    predictions[inactive_saturation] = 1 - np.exp(-alpha * predictions[inactive_saturation] - beta)
+
+    return 1 - predictions
 
 
 def predict_on_target(transcript_seq: str, model: tf.keras.Model):