Mining the Biomedical Literature Pengyu Hong State of the Art and Challenges Acknowledgement: Dr....

Mining the Biomedical Literature

Pengyu Hong

State of the Art and Challenges

Acknowledgement: Dr. Hagit Shatkay

http://www.shatkay.org/

2

Literature Experiments

e.g., Pathway

+ New Findings

e.g., Microarray

3

gene symptom disease

• Type 2 diabetes• Hypercholesterolemia• Coronary artery

disease• Essential hypertension

Insulin ResistanceTNFRSF1B

Genetic Basis of Disease

4

• Information Extraction• Information Retrieval• Cases• Integrate Text with Other Type of

Data• Conclusion

Overview

5

Automated identification of certain kinds of facts from text.

Information Extraction

Example:

Phosphorylation

Kinase Protein

glycogen synthase kinase-3 (GSK-3) BIN2

BES1

“…BES1 is phosphorylated and appears to be destabilized by the glycogen synthase kinase-3 (GSK-3) BIN2…”

6


• Identify the relevant sentences

• Parse to extract specific information

• Assume “well-behaved” fact sentences

• Using co-occurrence relationships alone does not require parsing or good fact-structure

Usually it requires

7

Automatically retrieve documents satisfying the information need (Query) from a lot of documents

Information Retrieval

8

Means:

• Boolean query, Index based (e.g. “Gene and CDC”)


CD

Chagas' disease

cytosine deaminase

Crohn‘s disease

capillary density

Cortical dysplasia

(54,745 Pubmed entries)

compact disk...

Synonymy (AGP1, aka, Amino Acid Permease1)

Polysemy

9

Means:

• Boolean query, Index based

• Similarity query, e.g., Vector based.


Semantic search

TIME (Sept 5, 2005): Search engines are good at matching words … The next step is semantic search – looking for meaning, not just matching key words. … Nervana, which analyzes language by linking word patterns contextually to answer questions in defined subject areas, such as medical-research literature.

10

Both information extraction and information retrieval require to process text and store text and the processed results.

11

Text Processing

• Lexical Analysis: Tokenization, breaking the text into basic

building blocks. (Paragraphs, Sentences, Phrases, Words)

12

The growth-promoting effect of growth hormone (GH) is primarily mediated by insulin-like growth factor-1 (IGF-1). The liver is the main source of circulating IGF-I. The authors have used rodent primary hepatocytes for studies on pharmacological intervention of IGF-I mRNA expression. A 96-well nonradioactive IGF-1 mRNA quantification assay was developed, based on the hybridization of sense and antisense RNA probes, to replicate membranes with crude hepatocyte lysates.

The sense hybridization was used as an internal standard. The antagonistic properties of a set of GH-receptor binding compounds were evaluated. Two compounds were found to down-regulate IGF-I mRNA. Effects due to metabolic inhibition or toxicity were excluded using a cell proliferation assay.

13

Text Processing



• Parsing/Processing: (Identify Noun-phrases, Verb-phrases,

names and/or other proper nouns, stem)

14

The growth-promoting effect of growth hormone (GH) is primarily mediated by insulin-like growth factor-1 (IGF-1). The liver is the main source of circulating IGF-I. The authors have used rodent primary hepatocytes for studies on pharmacological intervention of IGF-I mRNA expression. A 96-well nonradioactive IGF-1 mRNA quantification assay was developed, based on the hybridization of sense and antisense RNA probes, to replicate membranes with crude hepatocyte lysates.

The sense hybridization was used as an internal standard. The antagonistic properties of a set of GH-receptor binding compounds were evaluated. Two compounds were found to down-regulate IGF-I mRNA. Effects due to metabolic inhibition or toxicity were excluded using a cell proliferation assay.

studystem

15

Text Processing



• Parsing/Processing:

* Part-of-Speech (PoS) TaggingInitial assignment of syntactic categories to a sequence of words. Categories may include: noun, verb, determiner, adjective, adverb, proper noun, etc.

PoS Tagging Paradigms: • Rule based [Brill 92, Brill 99]• Stochastic [Charniak93, Allan95…]• Supervised vs. Unsupervised [Schütze 93], more…

16

Text Processing



• Parsing/Processing:

* Part-of-Speech (PoS) Tagging

* Parsing: Complete assignment of syntactic structure to a word sequence. Including: Noun phrase, Verb phrase, Prepositional phrase, etc.

• Full parsing: Full assignment of every word and phrase to a syntactic structure . [e.g. Grishman95]. Slow, and error-prone...• Shallow/Partial parsing: Partial assignment of syntax structures. “… parse conservatively – only if there is strong evidence of the correctness of a reduction.” [Grishman97 ] [Appelt et al. 95, Appelt & Israel 99, Feldman et al. 2000]

17

Text Storage

Support quick access to facts/documents:

Terms/words

Structure

A document

Query is not 1-Dimensional

18

Information Extraction - DetailsExtract what?

• Entities: e.g., genes, proteins, diseases, chemical compounds, etc.

• Relationships: e.g., phosphorylation, activation of a gene by a transcription factor, etc.

• Events: e.g., a protein is activated, a gene is transcribed, etc.

It is hard!• Entities can have synonyms and be referred as anaphora (e.g.,

this gene, that protein, the former, etc.). Unrelated entities may share a name (polysyms).

• Relationships and events can be stated in various styles and indirect ways.

19

Information Extraction – Details (cont.)

Rule-BasedExplicit rules for identifying entities and relations are defined.

Kinase_Related_Activity :- <regulate> <protein> <kinase> <activity>|<activate> <protein> <kinase> |<increase> <rate> <protein> <kinase> <activity>;

Rules can be:

• Hand-coded [e.g. Appelt et al. 95, Hobbs et al. 97]

• Automatically acquired from examples (Supervised learning) [e.g. Riloff 93, Califf 98, Soderland 99]

• Automatically acquired, no examples (Unsupervised learning) [e.g. Yangarber et al. 2000]

20

Information Extraction – Details (cont.)

Statistical Machine-LearningLearn probabilistic models for “typical” sequences denoting entities and relations.

[Bikel et al. 97, Leek 97, Seymore et. al. 99, Ray and Craven 2001]

For every name-class, C, a bi-gram, Markov probability Pc(Wi|Wi-1).

Induces: P(C|W1…Wn) P(W1…Wn|C) = Pc(Wi|Wi-1)

High-level, transition and emission probabilities: P(C| C-1, W-1), P(W | C, W-1 )

start action entity end

Example:

{regulate, activate, increase

……}

{MAPK, MAPKK, MAPKKK

……}

property

{extent, rate,

frequency

……}

21

DB: Database of documents.

Vocabulary: {t1,…,tM } (Terms in DB, produced by the tokenization stage)

Index Structure: A term all the documents containing it.

Information Retrieval – DetailsBoolean Queries

acquired immunodeficiency

asthma

blood

blood pressure

IndexDatabase

22

DB: Database of documents.

Vocabulary: {v1,…,vM } {Terms in DB}

Document dDB: Vector, <w1d,…,wM

d>, of weights.

Information Retrieval – DetailsThe Vector Model

Weighting Principles • Document frequency: Terms occurring in a few documents are

more useful than terms occurring in many.

• Local term frequency: Terms occurring frequently within a document are likely to be significant for the document.

• Document length: A term occurring the same # of times in a long document and in a short one has less significance in the long one.

• Relevance: Terms occurring in documents judged as relevant to a query, are likely to be significant (WRT the query).

[Sparck Jones et al. 98]

23

Some Weighting Schemes:

Binary

TF Wid = fi

d = # of times ti occurs in d.

Wid=

fid

fi

(fi= # of docs containing ti)TF X IDF(one version...)

Information Retrieval - Details(cont.)

Wid =

1 if ti d

0 otherwise

Consider Local term frequency

Consider Local term frequency and Document frequency

24

Document d= <w1d,…,wM

d>DB

Query q = < w1q,…,wM

q> (q could itself be a document in DB...)


Vector-Based similarity

Sim(q, d) = cosine (q, d )

= q • d

|q| |d|

d

q

[Salton89, Witten et al99] Introductory IR.

25

Information Retrieval - Details (cont.)

[Sparck Jones et al. 98, Sahami98, Ponte&Croft 98, Hoffman 99]

Probabilistic Models

Query q ; Document d

Log[Log[PP(relevant | (relevant | dd, , qq))

PP(Irrelevant | (Irrelevant | dd, , qq))]]Maximize log-odds:Maximize log-odds:

• Goal:Goal: Find all Find all dd’s such that ’s such that PP(relevant | (relevant | dd, , qq) is high) is high

26


Latent Semantics Analysis [Dumais, Deerwester et al,1988,1990]

Motivation: Overcoming synonymy and polysemy. Reducing dimensionality.

Idea: Project from “explicit term” space to a lower dimension, “abstract concept” space.

Methodology: Singular Value Decomposition (SVD) applied to the document-term matrix. Highest singular values are used as the features for

representing documents.

27

Information Retrieval- Details(cont.)

Text Categorization (semantic)

Automatically place documents in right categories so as to make them easy-to-find.

......

Cancer

Apoptosis Elongation

28

Information Retrieval-Details(cont.)

Rule-Based Text Classification

A knowledge-engineering approach.

Boolean rules (DNF), based on the presence/absence of specific terms within the document, decide its membership in the class. (e.g. the CONSTRUE system [Hayes et al. 90,92] )

Example: If ( (<GENE_Name> ⋀ transcript) ⋁ ((<GENE_Name> Western Blot) ⋀ ⋁ ((<GENE_Name> Northern Blot))⋀ Then GeneExpressionDoc Else Gene⌝ ExpressionDoc

29


Machine Learning for Text Classification (supervised)

• Take a training set of pre-classified documents

• Build a model for the classes from the training examples

• Assign each new document to the class that best fits it (e.g. closest or most-probable class.)

Types of class assignment:

Hard: Each document belongs to exactly one class

Soft: Each document is assigned a “degree of membership” in several classes

30


Machine Learning for Text Classification (supervised)

* Training set: pre-classified documents

* Learning task: Infer C from the data

...< , C( )>; < , C( )>; < , C( )>;T=

31

Examples of Supervised Learning methods

• Memorize the classes of all documents• Choose a distance/similarity function between documents• Assign a new document to the class of the closest document

Nearest Neighbor

Main drawback: Does not summarize the data; classification

of a new document can be excruciatingly slow, comparing it against ALL documents in the dataset.

32

Machine Learning Methods for Text Classification


Builds a class term-weight vector, , summarizing the document vectors in the class (and those not in it). Assigns new document vector , to class

C

d

)],([argmax dCcosc

C

Rocchio (1971)

[Rocchio71, Buckley et al. 94, Lewis et al. 96, Joachims97, Schapire et al. 98]

33

Naïve Bayes Classifier

• Assumes conditional independence among features, given the class

• Builds for each class, C, a term distribution model, CM

• Assigns new document d’ to class C’ s.t.

Using Bayes’ rule and the “Naïve” assumption:

CM : <P(t1d|d C), P(t2d|d C),…, P(tnd|d C)>

(where: d – document, ti –term)

P(CM | d’) = P(d’|CM) P(CM )

P( d’ )= kP(d’|CM)= k P(tid|d CM )∏

ti d’

[McCallum& Nigam 99]

)]|'([argmax' CC

M

MdPC

34

Machine Learning Methods for Text Classification (cont.)


[Joachims 98, Yang and Liu 99, Yang 2001, Schapire et al.98]

Partitions the training documents by a maximum-margin hyperplane.Classifies new document-vector, , based on its position WRT the separating hyperplane.

d

Support Vector Machine

Boosting/Mixtures

Combines multiple weak classifiers.

35

Evaluating Extraction and Retrieval

To say how good a system is we need:

1. Performance metrics (numerical measures)

2. Benchmarks, on which performance is measured (the gold-standard).

36

Evaluating Extraction and Retrieval(cont.)

Performance Metrics

N items (e.g. documents, terms or sentences) in the collection

REL: Relevant items (documents, terms or sentences) in the collection.These SHOULD be extracted or retrieved.

RETR: Retrieved items (e.g. documents, terms or sentences) are actually extracted/retrieved

Some correctly (A = |REL ⋀ RETR|),Some incorrectly (B = |RETR – REL| )

|RETR| = A+B

37

Evaluating Extraction and Retrieval(cont.)

Performance Metrics (cont.)

|RETR – REL| = B

Collection

REL RETR

|REL RETR| = ⋀ A

|Collection| = N

|REL-RETR| = D

|NotREL – RETR| = C

38


Precision: P = A/(A+B) How many of the retrieved/extracted items are correct

Recall: R = A/(A+D)How many of the items that should be retrieved are recovered

Accuracy: (A+C)/N (Ratio of Correctly classified items)

F-score: 2PR / (P+R)Harmonic mean, in the range [0,1]

Combination Scores:

Fβ-score: (1+β2)PR / (β2·P + R)

β >1 Prefer recall, β <1 Prefer precision

E-measure: 1 – F(β)-scoreInversely proportional to performance (Error measure).

39


Precision-Recall Curves

4 relevant documents in the collection.

7 retrieved and ranked.

1

7

6

5

4

3

225% Recall

50%

75%

100%

6675

66

100

0102030405060708090

100

0 25 50 75 100

Recall

Pre

cisi

on

40


Average ScoresAverage Precision: Average the precision over all the ranks in which a relevant document is retrieved.

Mean Average Precision: Mean of the Average Precision over all the queries.

Micro-Average: Average over individual items across queries

Macro-Average: Average over queries

For a given rank n, Pn: Precision at rank n (P@n)

R-Precision: PR where R is the number of relevant documents

Accounting for Ranks

41

Overview


Data• Conclusion

42

T. Leek, MSc thesis, 1997 [Leek97]


•Gene localization on chromosomes.

Example: Gene: HKE4 located on Chromosome: 6.

•HMMs characterize sentences describing localization relationships.

•Gene names and Chromosomes are identified through heuristics.

•Words denoting location (e.g. mapped, located, discovered) and methods (e.g. southern, blot) are pre-defined.

•Trained and tested on sentences from OMIM abstracts.

•Scored based on correct population of the relation slots.

method1

start

method2verb1 verb2Gene Chrom

end

“Southern analysis shows that HKE4 is located on human chromosome 6.”

43

M. Craven et al [Craven&Kumlien99, Ray&Craven01, Skounakis,Craven&Ray03]

•Protein sub-cellular localization and gene-disorder associations.

Examples: Protein: Enzyme UBC6 localized to Endoplasmic Reticulum.

Gene: PSEN1 associated with Alzheimer Disease

•HMMs (and their extensions) characterize sentences describing sub-cellular localization, and disease association.

(Other models for sub-cellular localization in [Craven&Kumlien99])

•HMMs’ states represent structural segments (e.g. NP_segment)

•Training: Sentences whose gene/protein/location/disease words are tagged, based on information from YPD and OMIM.(Protein and localization lexicon was provided in [Craven&Kumlien99])

•Scored on correct classification of relevant/irrelevant sentences.

Information Extraction (cont.)

44


•Protein-protein interactions.

Example: Protein: Spatzle Activates Protein: Toll

•Based on co-occurrence of the form “... p1...I1...p2...” within a sentence, where p1, p2 are proteins and I1 is an interaction term.

•Protein names and interaction terms (e.g. activate, bind, inhibit)are provided as a “dictionary”.

•Does not use formal modeling or machine-learning.

•Applied to two systems in Drosophila:

The Pelle system (6 proteins, 9 interactions) and

The cell-cycle control system (91 proteins), without quantitative analysis of the results.

A. Blaschke, A. Valencia et al,1999

45

•Gene-gene relations.

Example:

•Based on co-occurrence of the form “... g1...g2...” within a Pubmed abstract, where g1, g2 are gene names.

•Gene names are provided as a “dictionary”, harvested from HUGO, LocusLink, and other sources.

•Does not use formal modeling or machine-learning.

•Applied to 13,712 named human genes and millions of PubMed abstracts (Most extensive!)

•No extensive quantitative results analysis.


T. Jenssen, E. Hovig et al, 2001 (PubGene)

NR4A3NR4A2

[Pearson01] Discussion of this approach.

46

•Fukuda et al: PROPER. Rule-based (hand-coded); Identifies protein and domain names, without a dictionary. Tested on two sets of 30 and 50 domain-specific abstracts; 93-98% recall, 91-94% precision.

•Krauthammer et al: Uses gene/protein name dictionary, extracted from GenBank;Encodes text and dictionary as nucleotide sequences, and matches text to the dictionary using BLAST.

Tested on 1, full-text, review article (1162 gene/protein occurrences);78.8% recall, 71.7% precision.


Extracting Gene and Protein Names [Fukuda et al 98, Krauthammer et al 00, Tanabe&Wilbur 02, Hanisch et al 03]

47

•Tanabe&Wilbur: ABGene. Retrain Brill POS tagger, (on 7000 hand- tagged sentences from Medline abstracts), to identify a

new tag Gene. Augment with rules to eliminate false positives and recover false- negatives. Use a Naïve Bayes classifier (IR), to rank documents by likelihood to contain genes/proteins.

Tested on 56,469 Medline abstracts;Results for low-scoring documents, Precision <66% Recall <60%;for high-scoring gene-documents: Precision >85%, Recall >70%.

•Hanisch et al: Built a dictionary of ~38,000 proteins, and ~152,000 synonyms, combining multiple resources (OMIM, Swissprot, HUGO) and pruning/cleaning process. Extraction uses the dictionary and a scoring+optimization method to do better than naïve pattern matching.

Tested over 470 abstracts with 141 proteins occurring in them;Best case recall 90%, precision: 95%.

Extracting Gene and Protein Names (Cont.)

[Schwartz&Hearst PSB03, BioCreative04, Mika & Rost ISMB04]

48


Fish Oil

Blood Viscosity

Platelet aggregability

Vascular Reactivity Reduces(and co-occurs)

Raynaud’s Syndrome

Increased(and co-occurs)

Fish OilRaynaud’s Syndrome

•Based on transitivity of relationships in co-occurrence graph.

•This idea can be used to discover new facts by co-occurrence

[Swanson 86,Swanson87,Swanson90, Swanson and Smalheiser99, Weeber et al. 2001, Stapley & Benoit 2000, Srinivasan 2003, Srivinasan 2004]

Can Reduce

49

Information Retrieval (cont.)

Finding Protein-Protein-Interaction Abstracts [Marcotte et al., 01]

Yeast Medline(YM)665,807 abstracts

260 Protein-Protein Interaction (PPI) abstracts

Find Discriminating Words (DW) Pr(wA| AYM) != Pr(wA| A PPI)

A

Abstract

Naïve Bayes Classifier

APPI

APPI

Reaches about 65% precision at 65% recall on a set of 325 abstracts.

Train

50

Predicting Protein Subcellular Location[Stapley et al. 02]

Information Retrieval (cont.)

http://www.nobel.se/medicine/laureates/1999/press.html

Genes/proteins of various organelles

...

AbstractsMentioning Genes.

......

<w11,…,w1M>

<w11,…,w2M>

<wN1,…,wNM>

...

Term-vector Representation

TrainClassifiers

SVM

<wNew1,…,wNewM>

Organelle for <wNew1,…,wNewM>

Mitochondria genes/proteins

Nucleus genes/proteins

Cell Membrane genes/proteins

51

Overview


Data• Conclusion

52

Protein Homology by Sequence and Text [Chang et al. 01]

Integrate Text with Other Type of Data

P1 V G A H …P2

V G A N …P3

V E A D …P4

V N A N …

Find sequence homologyusing Psi-Blast, multiplealignment.

T2

T3

T4

T1

Create“documents” from all text associated with each protein (annotations and abstracts)

Find text similarity

Discard alignments with low document similarities.Create profile.

53

Gene Clustering by Expression and Text [Glenisson et al. 03,04]

Integrate Text with Other Type of Data

Find pair-wise dissimilarity between expression profiles

T2

T3

Tn

T1

Find pair-wise dissimilarity in text

Combine into a single gene-dissimilarity measure

G1

G3

G2

Gn

...

Create text-vectors based on all the abstracts curated for the genes (in SGD)

Hierarchically cluster the genes

......

54

Conclusion

Information Retrieval:• Looks for relevant documents• Does not give a tidy fact statement (Coarse granularity)• Can find relations among documents or document collections• Can create a coherent context for performing Information Extraction• Can foreshadow putative, yet-undiscovered relationships• Less sensitive to vocabulary and terminology

Information Extraction:• Extracts well-defined facts from the literature• Requires domain vocabulary or rules to identify these facts• Finds explicitly stated facts• Looks for facts stated within a sentence, a paragraph or a

single document (Fine granularity)

55

Conclusion

• Reduce dependency on vocabularies and nomenclature• Automate fact-finding about gene-disease interaction• Reconstruct metabolic, signaling or regulatory pathways• Integrate literature analysis with the analyses of large-

scale experimental data. A combined approach is likely to get us closer to our goal

Challenges:

Date post:	19-Dec-2015
Category:	Documents
View:	218 times
Download:	1 times

Mining the Biomedical Literature Pengyu Hong State of the Art and Challenges Acknowledgement: Dr....

Documents