1. Multi­way Relation Classification:
Application to Protein­Protein Interactions
Barbara Rosario
Marti A. Hearst
2005
Farzaneh Sarafraz
30 April 2009

2. HIV­1 Human Protein Interactions
Database
Pair of Proteins

Interaction type(s) between them

PubMed ID

(etc.)


3. Data
2224 records (now 5134)

65 interaction types (now 68)

809 proteins (now 1434 + 9 and 2295 pairs)

984 articles (now 3099)

Average 1.9 interactions per PP (max = 23)

Average 5.9 interactions per article (max = 90)


4. Goal
For every “triple”

PP
−
A (Article with unique pmid)
−
Find the interaction type

(ignore 7.7% of the triples with > 1 interaction)
−

5. NER
LocusLink

“Conservative” approach

No co­reference analysis

Not good recall

High precision


6. Method – assuming one interaction
For a subset of all the PPs (45%)

Get all full text articles
−
Get the sentences that have both PP
−
Group as “papers”
−
Also for a tripe PPA

Get the papers that cite A
−
Get the sentences that have PP and mention A
−
Group as “citances”
−

7. Training Data Construction
“papers”

0.5 sentence per triple (max 79)
−
50.6 sentences per interaction type (max 119)
−
“citances”

0.4 sentence per triple (max 105)
−
49.2 sentences per interaction type (max 162)
−
Include an interaction type if >40 in both


8. Interaction Types

9. Task
Given a PPA triple

Extract sentences that have PP

Predict for the entire PPA one of 10 interaction

types

10. Models
Generative

Dynamic Graphical Model
−
Simple Naïve Bayes Classifier
−
Discriminative

Neural Network (feed­forward, conjugate gradient)
−

11. Dynamic Graphical Model
Based on previous work

Roles: PROTEIN, NULL

Features: words


12. DM – Assumptions
There is an interaction

Single interaction per sentence

As many role states as words

Words = features

One feature node per role
−
Roles are hidden
−
Protein names may be masked
−

13. Evaluation
Document­level

(Not all the sentences describe an interaction)
−
For every triple an interaction is assigned to the
−
whole document
Using two methods:
−
Mj

Cf


14. Mj
for each triple

for each sentence of the triple
−
find the interaction that maximises the posterior

probability of the interaction given features
assign to all sentences of the triple the most

frequent interaction

15. Cf
get all conditional probabilities (do not assign

per sentence)
for each triple

choose the interaction that maximises the sum over
−
all the triple's sentences

16. Results

17. Comparison
Trigger word

70 triggers for 10 interactions
−
Co­occurrence
−
Choose the “most specific” type
−
If both specific or no trigger, choose nothing
−
Back­off: if in doubt, choose the most frequent
−
interaction

18. Comparison

Key(B): trigger word (backoff)
−
Base: the most frequent interaction
−

19. Sentence­Level Experiments
Manual annotation of 2114 sentences

68.3% disagreed with HIV database

Contacted some of the authors

DB error
−
Contradiction
−
“require” but under certain conditions “inhibit”


20. Sentence­Level Evaluation

21. Thank you.