Predicting Protein Function Using Machine-Learned Hierarchical Classifiers

1
Predicting Protein Function Using Machine-Learned
Hierarchical Classifiers

• Roman Eisner
• Supervisors Duane Szafron and Paul Lu

2
Outline

• Introduction
• Predictors
• Evaluation in a Hierarchy
• Local Predictor Design
• Experimental Results
• Conclusion

3
(No Transcript)
4
Proteins

• Functional Units in the cell
• Perform a Variety of Functions
• e.g. Catalysis of reactions, Structural and
  mechanical roles, transport of other molecules
• Can take years to study a single protein
• Any good leads would be helpful!

5
Protein Function Prediction and Protein Function
Determination

• Prediction
• An estimate of what function a protein performs
• Determination
• Work in a laboratory to observe and discover what
  function a protein performs
• Prediction complements determination

6
Proteins

• Chain of amino acids
• 20 Amino Acids
• FastA Format

gtP18077 R35A_HUMAN MSGRLWSKAIFAGYKRGLRNQREHTALLK
IEGVYARDETEFYLGKR CAYVYKAKNNTVTPGGKPNKTRVIWGKVTRAH
GNSGMVRAKFRSNL PAKAIGHRIRVMLYPSRI
7
Ontologies

• Standardized Vocabularies (Common Language)
• In biological literature, different terms can be
  used to describe the same function
• e.g. peroxiredoxin activity and
• thioredoxin peroxidase activity
• Can be structured in a hierarchy to show
  relationships

8
Gene Ontology

• Directed Acyclic Graph (DAG)
• Always changing
• Describes 3 aspects of protein annotations
• Molecular Function
• Biological Process
• Cellular Component

9
Gene Ontology

• Directed Acyclic Graph (DAG)
• Always changing
• Describes 3 aspects of protein annotations
• Molecular Function
• Biological Process
• Cellular Component

10
Hierarchical Ontologies

• Can help to represent a large number of classes
• Represent General and Specific data
• Some data is incomplete could become more
  specific in the future

11
Incomplete Annotations
12
Goal

• To predict the function of proteins given their
  sequence

13
Data Set

• Protein Sequences
• UniProt database
• Ontology
• Gene Ontology Molecular Function aspect
• Experimental Annotations
• Gene Ontology Annotation project _at_ EBI
• Pruned Ontology 406 nodes (out of 7,399) with
  20 proteins
• Final Data Set 14,362 proteins

14
Outline

• Introduction
• Predictors
• Evaluation in a Hierarchy
• Local Predictor Design
• Experimental Results
• Conclusion

15
Predictors

• Global
• BLAST NN
• Local
• PA-SVM
• PFAM-SVM
• Probabilistic Suffix Trees

16
Predictors

• Global
• BLAST NN
• Local
• PA-SVM
• PFAM-SVM
• Probabilistic Suffix Trees

• Linear

17
Why Linear SVMs?

• Accurate
• Explainability
• Each term in the dot product in meaningful

18
PA-SVM

• Proteome Analyst

19
PFAM-SVM

• Hidden Markov Models

20
PST

• Probabilistic Suffix Trees
• Efficient Markov chains
• Model the protein sequences directly
• Prediction

21
BLAST

• Protein Sequence Alignment for a query protein
  against any set of protein sequences

22
BLAST
23
Outline

• Introduction
• Predictors
• Evaluation in a Hierarchy
• Local Predictor Design
• Experimental Results
• Conclusion

24
Evaluating Predictions in a Hierarchy

• Not all errors are equivalent
• Error to sibling different than error to
  unrelated part of hierarchy
• Proteins can perform more than one function
• Need to combine predictions of multiple functions
  into a single measure

25
Evaluating Predictions in a Hierarchy

• Semantics of the hierarchy True Path Rule
• Protein labeled with
• T -gt T, A1, A2
• Predicted functions
• S -gt S, A1, A2
• Precision 2/3 67
• Recall 2/3 67

26
Evaluating Predictions in a Hierarchy

• Protein labelled with
• T -gt T, A1, A2
• Predicted
• C1 -gt C1, T, A1, A2
• Precision 3/4 75
• Recall 3/3 100

27
Supervised Learning
28
Cross-Validation

• Used to estimate performance of classification
  system on future data
• 5 Fold Cross-Validation

29
Outline

• Introduction
• Predictors
• Evaluation in a Hierarchy
• Local Predictor Design
• Experimental Results
• Conclusion

30
Inclusive vs Exclusive Local Predictors

• In a system of local predictors, how should each
  local predictor behave?
• Two extremes
• A local predictor predicts positive only for
  those proteins that belong exactly at that node
• A local predictor predicts positive for those
  proteins that belong at or below them in the
  hierarchy
• No a priori reason to choose either

31
Exclusive Local Predictors
32
Inclusive Local Predictors
33
Training Set Design

• Proteins in the current folds training set can
  be used in any way
• Need to select for each local predictor
• Positive training examples
• Negative training examples

34
Training Set Design
35
Training Set Design
36
Training Set Design
37
Training Set Design
38
Training Set Design
39
Comparing Training Set Design Schemes

• Using PA-SVM

40
Exclusive have more exceptions
41
Lowering the Cost of Local Predictors

• Top-Down
• Compute local predictors top to bottom until a
  negative prediction is reached

42
Lowering the Cost of Local Predictors

• Top-Down
• Compute local predictors top to bottom until a
  negative prediction is reached

43
Lowering the Cost of Local Predictors

• Top-Down
• Compute local predictors top to bottom until a
  negative prediction is reached

44
Top-Down Search
45
Outline

• Introduction
• Predictors
• Evaluation in a Hierarchy
• Local Predictor Design
• Experimental Results
• Conclusion

46
Predictor Results
47
Similar and Dissimilar Proteins

• 89 of proteins at least one good BLAST hit
• Proteins which are similar (often homologous) to
  the set of well studied proteins
• 11 of proteins no good BLAST hit
• Proteins which are not similar to the set of well
  studied proteins

48
Coverage

• Coverage Percentage of proteins for which a
  prediction is made

49
Similar Proteins Exploiting BLAST

• BLAST is fast and accurate when a good hit is
  found
• Can exploit this to lower the cost of local
  predictors
• Generate candidate nodes
• Only compute local predictors for candidate nodes
• Candidate node set should have
• High Recall
• Minimal Size

50
Similar Proteins Exploiting BLAST

• candidate nodes generating methods
• Searching outward from BLAST hit
• Performing the union of more than one BLAST hits
  annotations

51
Similar Proteins Exploiting BLAST
52
Dissimilar Proteins

• The more interesting case

53
Comparison to Protfun

• On a pruned ontology (9 Gene Ontology classes)
• On 1,637 no good BLAST hit proteins

54
Future Work

• Try other two ontologies biological process and
  cellular component
• Use other local predictors
• More parameter tuning
• Predictor cost

55
Conclusion

• Protein Function Prediction provides good leads
  for Protein Function Determination
• Hierarchical ontologies can represent incomplete
  data allowing the prediction of more functions
• Considering the hierarchy
• More accurate Less Computationally Intensive
• Methods presented have a higher coverage than
  BLAST alone
• Results accepted to IEEE CIBCB 2005

56
Thanks to

• Duane Szafron and Paul Lu
• Brett Poulin and Russ Greiner
• Everyone in the Proteome Analyst research group

57
Incomplete Data Prediction

• Inclusive avoids using ambiguous (incomplete)
  training data
• Does this help?
• To test
• Train on more Incomplete Data
• Choose X of proteins, and move one annotation up
• Evaluation Predictions on Complete data

58
Robustness to Incomplete Data
59
Local vs Global Cross-Validation

• Some node predictors have as little as 20
  positive examples
• How to do cross-validation to make sure each
  predictor has enough positive training examples?

60
Local vs Global Cross-Validation

• Local cross-validation is invalid
• Predictions must be consistent
• Need fold isolation
• A single global split
• global cross-validation