Técnicas de IA para Biologia

10 - Usage of Gene Ontology

Usage of Gene Ontology

Inference in the Gene Ontology

Inference in the Gene Ontology

True Path Rule

• Gene Products are annotated to the most specific GO term
• Annotations are (implicitly) propagated to ancestor terms ($\sf is\_a$) as well as via $\sf part\_of$
• True Path Rule: path from a child term through all ancestors back to the root must be biologically accurrate

Inference in the Gene Ontology

Problem Example

• For one species (flies): chitin metabolism a child of cuticle synthesis
• But chitin metabolism also part of cell wall organization in yeast
• Yeast gene annotated with chitin biosynthesis implies annotation to cuticle biosynthesis, but yeast does not have cuticles

Inference in the Gene Ontology

Problem fix

• Introduction of new GO terms separating the two kinds of processes

Inference in the Gene Ontology

Inferences over GO Edges

• A number of inference rules have been established
• For instance:
$$A \stackrel{is\_a}{\longrightarrow} B \wedge B \stackrel{part\_of}{\longrightarrow} C \Rightarrow A \stackrel{part\_of}{\longrightarrow} C $$

Inference in the Gene Ontology

Example

[Term]
id: GO:0044444
name: cytoplasmic part
is_a: GO:0044424 ! intracellular part
relationship: part_of GO:0005737 ! cytoplasm

[Term]
id: GO:0005737
name: cytoplasm

[Term]
id: GO:0005739
name: mitochondrion
is_a: GO:0044444 ! cytoplasmic part

• We can infer that mitochondrion $\sf part\_of$ cytoplasm holds

Inference in the Gene Ontology

Example

• We can infer that mitochondrion $\sf part\_of$ cytoplasm holds

• This is not the same as saying mitochondrion $\sf is\_a$ cytoplasm

Inference in the Gene Ontology

Other Inference rules

• Transitivity
$$A \stackrel{is\_a}{\longrightarrow} B \wedge B \stackrel{is\_a}{\longrightarrow} C \Rightarrow A \stackrel{is\_a}{\longrightarrow} C $$ $$A \stackrel{part\_of}{\longrightarrow} B \wedge B \stackrel{part\_of}{\longrightarrow} C \Rightarrow A \stackrel{part\_of}{\longrightarrow} C $$
• Similar rules apply to $\sf has\_part$
• As well as to $\sf regulates$ and its subproperties $\sf positively\_regulates$ and $\sf negatively\_regulates$

Inference in the Gene Ontology

Intra-GO Cross-Product Definitions

[Term]
id: GO:0000152
name: nuclear ubiquitin ligase complex
namespace: cellular_component
def: "A ubiquitin ligase complex found in the nucleus." [GOC:mah]
is_a: GO:0000151 ! ubiquitin ligase complex
is_a: GO:0044428 ! nuclear part
intersection_of: GO:0000151 ! ubiquitin ligase complex
intersection_of: part_of GO:0005634 ! nucleus

• Intersections provide an equivalent definition
$$ GO:0000152 \equiv GO:0000151 \sqcap \exists part\_of.GO:0005634 $$
• Follows the design pattern of a more specific class $X$ (of general class $G$) differentiated by an additional discriminant $D$
• X is a G such that D
• GO:0000152 is a GO:0000151 such that $\exists part\_of.GO:0005634$

Inference in the Gene Ontology

External Cross-Product Definitions

[Term]
id: GO:0001510 ! RNA methylation
intersection_of: GO:0008152 ! metabolic process
intersection_of: OBO_REL:results_in_addition_of CHEBI:32875 !methyl group
intersection_of: OBO_REL:results_in_addition_to CHEBI:33697 !ribonucleic acid

• With external terms from the ChEBI ontology (Chemical Entities of Biological Interest)

Inference in the Gene Ontology

Reasoning with Cross-Product Definitions

[Term]
id: GO:0030223 ! neutrophil differentiation
intersection_of: GO:0030154 ! cell differentiation
intersection_of: OBO_REL:results_in_acquisition_of_features_of\ CL:0000775
   ! neutrophil

[Term]
id: GO:0030851 ! granulocyte differentiation
intersection_of: GO:0030154 ! cell differentiation
intersection_of: OBO_REL:results_in_acquisition_of_features_of\ CL:0000094
   ! granulocyte

[Term]
id: CL:0000775
name: neutrophil
is_a: CL:0000094 ! granulocyte

• We can infer that neutrophil differentiation is a subclass of granulocyte differentiation

Usage of Gene Ontology

Overrepresentation Analysis

Overrepresentation Analysis

Overview

• High-throughput technologies in molecular biology
• Allow to measure all genes in the genome experimentally
• DNA microarrays with thousands of probes
• Quantify the amount of corresponding sequences in the sample
• Typical microarray experiments:
• Compare gene expression profiles (their concentrations) under two or more biological conditions
• E.g., comparison between healthy and diseased tissue or different developmental stages
• Several replicate microarray experiments for each biological condition
• Statistical analysis for significant differences for each gene
• Outcome often a list of hundreds/thousands of differentially expressed genes

Overrepresentation Analysis

Idea

• Question: One or more specific GO term annotates more of the differentially expressed genes than one would expect by chance?
• Example
• Say 221 of 6000 yeast genes (3.7%) represented on a microarray are annotated to the GO term sporulation
• If we perform some experiment and observe 100 differentially expressed genes, 3 or 4 should be annotated with sporulation, merely by chance
• Suppose that 35 of 100 are annotated to sporulation
• We conclude that sporulation is overrepresented among differentially expressed genes

Overrepresentation Analysis

Issue

• We may based on such observations develop hypotheses to justify the outcome
• Determine subsequent experiments to test the hypothesis
• Multiple overrepresented GO terms may be indentified, inflating the number of significantly overrepresented terms
• List of 50 to 100 GO terms is not helpful in principle for determining which of the terms is the most characteristic

Overrepresentation Analysis

Solution

• Several algorithms based on hypergeometric distribution and related concepts
• Some on a term for term basis
• Irrelevant terms are omitted upfront (filtered)
• Still often many corrolated terms in results - propagation rule - which one is the most suitable?
• Being annotated to a given GO term, also implies annotation to its ancestors
• Tests for overrepresentation of similar terms are not statistically independent

Overrepresentation Analysis

Solution

• Parent-child algorithms
• Takes the propagation rule into account for determining overrepresentation
• Topology-based algorithms
• Find the most specific overrepresented terms
• Other model-based approaches
• Rather than a term for term analysis, an optimization problem is created that associates a score to a set of GO terms, trying to find an optimal combination of GO terms that together best explain the observed pattern

Usage of Gene Ontology

Semantic Similarity

Semantic Similarity

Overview

• Concepts in ontologies are connected by semantic relations
• Measures of similiarity for terms can be defined based on that
• Used for
• Validating results of gene expression clustering
• Predicting molecular interactions
• Disease gene prioritization
• Clinical diagnostics

Semantic Similarity

Basic Idea - Information Content

• For ontologies of $\sf is\_a$ relations
• Define a propability function $p$ such that $p(C_i)$ is the probability of encountering an instance of class $C_i$
• Recall that $x$ $instance\_of$ $A$ and $A$ $is\_a$ $B$ implies $x$ $instance\_of$ $B$, i.e., $p$ is monotonically increasing as we move to more general concepts
• Unique root $C$ has $p(C)=1$
• Information Content of a term t:
$$ IC(t) = - log\; p(t) $$
• More general concepts provide less information

Semantic Similarity

Probability of a term

• Intrinsic: uses internal structure of the ontology
• e.g. number of descendants / total number of Entities
• Extrinsic: uses frequency on a dataset
• probability that a randomly chosen protein is annotated to $t$, if we choose the protein from the set of all proteins under consideration
• Terms that annotate many genes have low information content
• Terms that annotate few genes have high information content

Semantic Similarity

Example on Information Content

• Annotations on 1000 documents about carnivores

Semantic Similarity

Resnik Semantic Similarity

• The more information two terms share, the more similar they are
• Information shared between two terms is indicated by the information content of their Most Informative Common Ancestor (MICA)
• Similarity between two terms determined as follows:
$$sim(t_1,t_2) = IC(MICA(t_1,t_2)) = max_{t\in Anc(t_1)\cap Anc(t_2)} IC(t)$$

Semantic Similarity

Example on similarity

• Similarity between cheetah and lion

Semantic Similarity

Example on similarity

• Similarity between beagle and wildcat

Semantic Similarity

Improvements

• Variants of similarity measure have been developed
• Take also into account the distance between the terms
• Otherwise wolf and fox are are as similar as beagle and fox
• Can be measured by path length
• But depends on the ontology
• Maximum similarity is reached if the terms are identical
$$ sim_{Lin}(t_1,t_2) = (2 \times max_{t\in Anc(t_1)\cap Anc(t_2)} IC(t))/ (IC(t_1) + IC(t_2))$$
• Further notions have been defined

Semantic Similarity

Applied to GO

• Focus is on similarity between genes annotated by terms - not similarity between terms
• Similarity measures defined based on the similarity of its annotations
• Maximum value of all pairs
• Average
• Also graph-based approaches relying on counting edges
• Set-based measures, e.g., intersection of annotations/union of annotations

Usage of Gene Ontology

Summary

• Inference in the Gene Ontology
• Overrepresentation Analysis
• Semantic Similarity

Further reading:

• Robinson and Bauer, Introduction to Bio-Ontologies, Chapters 8, 10, 12
• Dessimoz and Skunca, The Gene Ontology Handbook
• GO webpage http://geneontology.org/

Exercises

Inference

[Term]
id: GO:0007519
name: skeletal muscle tissue development
is_a: GO:0014706 ! striated muscle tissue development
relationship: part_of GO:0060538 ! skeletal muscle organ\development

[Term]
id: GO:0014706
name: striated muscle tissue development
is_a: GO:0060537 ! muscle tissue development

[Term]
id: GO:0060537
name: muscle tissue development
is_a: GO:0009888 ! tissue development
relationship: part_of GO:0007517 ! muscle organ development

• The CSRPP3 gene is annotated to the GO term skeletal muscle tissue development. What other annotations can we infer for this protein and why?

Exercises

Inference

[Term]
id: GO:0006310
name: DNA recombination
is_a: GO:0006259 ! DNA metabolic process
[Term]
id: GO:0042148
name: strand invasion
is_a: GO:0006259 ! DNA metabolic process
relationship: part_of GO:0006310 ! DNA recombination
[Term]
id: GO:0060542
name: regulation of strand invasion
is_a: GO:0000018 ! regulation of DNA recombination
relationship: regulates GO:0042148 ! strand invasion
[Term]
id: GO:0060543
name: negative regulation of strand invasion
is_a: GO:0045910 ! negative regulation of DNA recombination
is_a: GO:0060542 ! regulation of strand invasion
relationship: negatively_regulates GO:0042148 ! strand invasion

• MPH1 protein is annotated to GO:0060543. What other ...

Exercises

Gene Ontology (Python)

• Download basic version of GO ontology
• http://current.geneontology.org/ontology/go-basic.obo
• Download GO Semantic Similarity file http://labs.rd.ciencias.ulisboa.pt/dishin/go202104.db.gz
• Install:
• pip install goatools ssmpy

Exercises

Gene Ontology (Python) - Basics

from goatools import obo_parser

go_obo = 'go-basic.obo'
# create a dictionary of the GO terms
go = obo_parser.GODag(go_obo)

go_id = 'GO:0048528'
go_term = go[go_id]
print(go_term)
print('GO term name: {}'.format(go_term.name))
print('GO term namespace: {}'.format(go_term.namespace))

for term in go_term.parents:
    print(term)   
for term in go_term.children:
    print(term)

rec = go[go_id]
parents = rec.get_all_parents()
children = rec.get_all_children()
for term in parents.union(children):
    print(go[term])
		

Exercises

Gene Ontology (Python) - common ancestors

• Find the nearest common ancestor of GO:0048527 and GO:0097178

def common_parent_go_ids(terms, go):
    # Find candidates from first
    rec = go[terms[0]]
    candidates = rec.get_all_parents()
    candidates.update({terms[0]})

    # Find intersection with second to nth term
    for term in terms[1:]:
        rec = go[term]
        parents = rec.get_all_parents()
        parents.update({term})

        # Find the intersection with the candidates, and update.
        candidates.intersection_update(parents)

    return candidates
		

Exercises

Gene Ontology (Python) - common ancestors

def deepest_common_ancestor(terms, go):
    # Take the element at maximum depth. 
    return max(common_parent_go_ids(terms, go), key=lambda t: go[t].depth)

...
go_id_id1_dca = deepest_common_ancestor([go_id, go_id1], go)
print('The nearest common ancestor of\n\t{} ({})\nand\n\t{} ({})\nis\n\t{} ({})'
      .format(go_id, go[go_id].name, 
             go_id1, go[go_id1].name,
             go_id_id1_dca, go[go_id_id1_dca].name))
		

Exercises

Gene Ontology (Python)

• What is the name of the term GO:0097192?
• What is the most specific term that is parent of both GO:0097191 and GO:0038034?

Exercises

Gene Ontology (Python) - semantic similarity

• Calculate the semantic similarity between GO:0048364 (root development) and GO:0048486 (parasympathetic nervous system development) based on the number of branches separating them.

def min_branch_length(go_id1, go_id2, go):
    # First get the deepest common ancestor
    dca = ...

    # Then get the distance from the DCA to each term
    dca_depth = go[dca].depth
    d1 = go[go_id1].depth - dca_depth
    d2 = go[go_id2].depth - dca_depth

    # Return the total distance - i.e., to the deepest common ancestor and back.
    return ...
		

Exercises

Gene Ontology (Python) - semantic similarity

def semantic_distance(go_id1, go_id2, go):
    return min_branch_length(go_id1, go_id2, go)

def semantic_similarity(go_id1, go_id2, go):
    return 1.0 / float(semantic_distance(go_id1, go_id2, go))

...
print('The semantic similarity between terms {} and {} is {}.'.format(
      go_id1, go_id2, sim))
		

Exercises

Gene Ontology (Python) - semantic similarity

• Using DiShIn: https://dishin.readthedocs.io/
• Download and uncompress: http://labs.rd.ciencias.ulisboa.pt/dishin/go202104.db.gz

import ssmpy
 ssmpy.semantic_base("go.db")
 ssmpy.ssm.intrinsic = True
 e1 = ssmpy.get_id(go_id1.replace(":", "_"))
 e2 = ssmpy.get_id(go_id2.replace(":", "_"))
 print(e1, e2)
 print(ssmpy.ssm_resnik(e1,e2))
 print(ssmpy.ssm_lin(e1,e2))
		

Exercises

Gene Ontology (Python) - semantic similarity

• Similarity between proteins

e1 = ssmpy.get_uniprot_annotations("Q12345")
e2 = ssmpy.get_uniprot_annotations("Q12346")
ssmpy.ssm_multiple(ssmpy.ssm_resnik, e1, e2)
ssmpy.ssm_multiple(ssmpy.ssm_lin, e1, e2)