Merge pull request #82 from kundajelab/bugfix

Reverting default setting, minor fixes

modisco.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: modisco
-Version: 0.5.10.0
+Version: 0.5.10.2
 Summary: TF MOtif Discovery from Importance SCOres
 Home-page: https://github.com/kundajelab/tfmodisco
 License: UNKNOWN

modisco/coordproducers.py

@@ -678,9 +678,12 @@ def identify_coords(score_track, pos_threshold, neg_threshold,
     # coordinates are identified
     cp_score_track = [np.array(x) for x in score_track]
     #if a position is less than the threshold, set it to -np.inf
+    #Note that the threshold comparisons need to be >= and not just > for
+    # cases where there are lots of ties at the high end (e.g. with an IR
+    # tranformation that gives a lot of values that have a precision of 1.0)
     cp_score_track = [
-        np.array([np.abs(y) if (y > pos_threshold
-                        or y < neg_threshold)
+        np.array([np.abs(y) if (y >= pos_threshold
+                        or y <= neg_threshold)
                        else -np.inf for y in x])
         for x in cp_score_track]
 
@@ -712,8 +715,8 @@ def identify_coords(score_track, pos_threshold, neg_threshold,
                     other_info = dict([
                      (track_name, track[example_idx][argmax])
                      for (track_name, track) in other_info_tracks.items()])) 
-                assert (coord.score > pos_threshold
-                        or coord.score < neg_threshold)
+                assert (coord.score >= pos_threshold
+                        or coord.score <= neg_threshold)
                 coords.append(coord)
             else:
                 assert False,\
@@ -797,6 +800,8 @@ def refine_thresholds_based_on_frac_passing(
                 a=np.abs(vals),
                 q=100*(1-max_passing_windows_frac)) 
             neg_threshold = -pos_threshold
+        if (verbose):
+            print("New thresholds are",pos_threshold,"and",neg_threshold)
 
     return pos_threshold, neg_threshold
 

modisco/tfmodisco_workflow/workflow.py

@@ -166,7 +166,7 @@ class TfModiscoWorkflow(object):
     def __init__(self,
                  seqlets_to_patterns_factory=
                  seqlets_to_patterns.TfModiscoSeqletsToPatternsFactory(),
-                 sliding_window_size=[5,9,13,17,21], flank_size=10,
+                 sliding_window_size=21, flank_size=10,
                  overlap_portion=0.5,
                  min_metacluster_size=100,
                  min_metacluster_size_frac=0.01,

modisco/util.py

@@ -474,7 +474,7 @@ def compute_per_position_ic(ppm, background, pseudocount):
     assert len(ppm.shape)==2
     assert ppm.shape[1]==len(background),\
             "Make sure the letter axis is the second axis"
-    if (np.max(np.abs(np.sum(ppm, axis=1)-1.0)) < 1e-5):
+    if (not np.allclose(np.sum(ppm, axis=1), 1.0, atol=1.0e-5)):
         print("WARNING: Probabilities don't sum to 1 in all the rows; this can"
               +" be caused by zero-padding. Will renormalize. PPM:\n"
               +str(ppm)

setup.py

@@ -6,7 +6,7 @@
           description='TF MOtif Discovery from Importance SCOres',
           long_description="""Algorithm for discovering consolidated patterns from base-pair-level importance scores""",
           url='https://github.com/kundajelab/tfmodisco',
-          version='0.5.10.1',
+          version='0.5.10.2',
           packages=find_packages(),
           package_data={
                 '': ['cluster/phenograph/louvain/*convert*', 'cluster/phenograph/louvain/*community*', 'cluster/phenograph/louvain/*hierarchy*']