v0.0.9

modiscolite/__init__.py

@@ -8,4 +8,4 @@
 from . import cluster
 from . import report
 
-__version__ = '0.0.8'
+__version__ = '0.0.9'

modiscolite/affinitymat.py

@@ -240,8 +240,12 @@ def tsne_probs_calc(self, distances_nn, neighbors_nn):
 		conditional_P = sklearn.manifold._utils._binary_search_perplexity(
 			distances, self.perplexity, verbose=False)
 
+		eps = 1e-8
+		marginal_sum = conditional_P.sum(axis=-1)
+		marginal_sum[marginal_sum < eps] = eps
+
 		#normalize the conditional_P to sum to 1 across the rows
-		conditional_P = conditional_P/np.sum(conditional_P, axis=-1)[:,None]
+		conditional_P = conditional_P / marginal_sum[:,None]
 
 		data = []
 		rows = []

modiscolite/aggregator.py

@@ -19,6 +19,9 @@ def polish_pattern(pattern, min_frac, min_num, track_set, flank, window_size, bg
 		track_set=track_set, left_flank_to_add=flank,
 		right_flank_to_add=flank)
 
+	if pattern is None:
+		return None
+
 	# Trim by IC
 	ppm = pattern.sequence
 	per_pos_ic = util.compute_per_position_ic(
@@ -382,13 +385,14 @@ def SimilarPatternsCollapser(patterns, track_set,
 					min_num=min_num, track_set=track_set, flank=flank_to_add, 
 					window_size=window_size, bg_freq=bg_freq)
 
-				for k in range(len(patterns)):
-					#Replace EVERY case where the parent or child
-					# pattern is present with the new pattern. This
-					# effectively does single-linkage.
-					if (patterns[k]==parent_pattern or
-						patterns[k]==child_pattern):
-						patterns[k]=new_pattern
+				if new_pattern is not None:
+					for k in range(len(patterns)):
+						#Replace EVERY case where the parent or child
+						# pattern is present with the new pattern. This
+						# effectively does single-linkage.
+						if (patterns[k]==parent_pattern or
+							patterns[k]==child_pattern):
+							patterns[k]=new_pattern
 
 		merge_occurred_last_iteration = (len(indices_to_merge) > 0)
 

modiscolite/core.py

@@ -140,7 +140,6 @@ def compute_subpatterns(self, perplexity, n_seeds, n_iterations=-1):
 		distmat_nn = np.log((1.0/(0.5*np.maximum(affmat_nn, 0.0000001)))-1)
 		distmat_nn = np.maximum(distmat_nn, 0.0) #eliminate tiny neg floats
 
-
 		distmat_sp = scipy.sparse.coo_matrix(
 				(np.concatenate(distmat_nn, axis=0),
 				 (np.array([i for i in range(len(seqlet_neighbors))
@@ -151,11 +150,8 @@ def compute_subpatterns(self, perplexity, n_seeds, n_iterations=-1):
 		distmat_sp.sort_indices()
 
 		#do density adaptation
-		density_adapted_affmat_transformer =\
-			affinitymat.NNTsneConditionalProbs(
-				perplexity=perplexity)
-		sp_density_adapted_affmat = density_adapted_affmat_transformer(
-										affmat_nn, seqlet_neighbors)
+		sp_density_adapted_affmat = affinitymat.NNTsneConditionalProbs(
+				perplexity=perplexity)(affmat_nn, seqlet_neighbors)
 
 		sp_density_adapted_affmat += sp_density_adapted_affmat.T
 		sp_density_adapted_affmat /= np.sum(sp_density_adapted_affmat.data)

modiscolite/tfmodisco.py

@@ -56,8 +56,6 @@ def _density_adaptation(affmat_nn, seqlet_neighbors, tsne_perplexity):
 
 	affmat_diags = scipy.sparse.diags(1.0 / normfactors)
 	affmat_nn += affmat_diags
-
-
 	return affmat_nn
 
 def _filter_patterns(patterns, min_seqlet_support, window_size, 
@@ -117,8 +115,9 @@ def _motif_from_clusters(seqlets, track_set, min_overlap,
 			min_num=min_num, track_set=track_set, flank=flank_to_add, 
 			window_size=window_size, bg_freq=bg_freq)
 
-		if np.sign(np.sum(pattern.contrib_scores)) == track_sign:
-			cluster_to_motif.append(pattern)
+		if pattern is not None:
+			if np.sign(np.sum(pattern.contrib_scores)) == track_sign:
+				cluster_to_motif.append(pattern)
 
 	return cluster_to_motif
 

setup.py

@@ -2,7 +2,7 @@
 
 setup(
 	name='modisco-lite',
-	version='0.0.8',
+	version='0.0.9',
 	author='Jacob Schreiber',
 	author_email='[email protected]',
 	packages=['modiscolite'],