diff options
Diffstat (limited to 'wqflask/base/trait.py')
-rw-r--r-- | wqflask/base/trait.py | 44 |
1 files changed, 42 insertions, 2 deletions
diff --git a/wqflask/base/trait.py b/wqflask/base/trait.py index 6c5ca8b2..af22b5a1 100644 --- a/wqflask/base/trait.py +++ b/wqflask/base/trait.py @@ -308,14 +308,54 @@ class GeneralTrait(object): if isinstance(trait_info[i], basestring): holder = unicode(trait_info[i], "utf8", "ignore") setattr(self, field, holder) +<<<<<<< HEAD + + description_string = unicode(str(self.description).strip(codecs.BOM_UTF8), 'utf-8') + target_string = unicode(str(self.probe_target_description).strip(codecs.BOM_UTF8), 'utf-8') + + if len(description_string) > 1 and description_string != 'None': + description_display = description_string + else: + description_display = self.symbol + + if (len(description_display) > 1 and description_display != 'N/A' and + len(target_string) > 1 and target_string != 'None'): + description_display = description_display + '; ' + target_string.strip() + + # Save it for the jinja2 template + self.description_display = description_display + + #XZ: trait_location_value is used for sorting + trait_location_repr = 'N/A' + trait_location_value = 1000000 + + if self.chr and self.mb: + #Checks if the chromosome number can be cast to an int (i.e. isn't "X" or "Y") + #This is so we can convert the location to a number used for sorting + trait_location_value = convert_location_to_value(self.chr, self.mb) + #try: + # trait_location_value = int(self.chr)*1000 + self.mb + #except ValueError: + # if self.chr.upper() == 'X': + # trait_location_value = 20*1000 + self.mb + # else: + # trait_location_value = (ord(str(self.chr).upper()[0])*1000 + + # self.mb) + + #ZS: Put this in function currently called "convert_location_to_value" + self.location_repr = 'Chr%s: %.6f' % (self.chr, float(self.mb)) + self.location_value = trait_location_value + +======= +>>>>>>> e0c5c1aae3aaaa1d81bcec36835a97e169dcc2e2 if self.dataset.type == 'Publish': self.confidential = 0 if self.pre_publication_description and not self.pubmed_id: self.confidential = 1 - - description = self.post_publication_description + description = self.post_publication_description + #If the dataset is confidential and the user has access to confidential #phenotype traits, then display the pre-publication description instead #of the post-publication description |