Learning to Extract Proteins and their Interactions from Medline Abstracts

Learning to Extract Proteins and their Interactions from Medline Abstracts Raymond J. Mooney Department of Computer Sciences Razvan Bunescu, Ruifang Ge, Rohit J. Kate, Yuk Wah Wong Edward M. Marcotte, Arun Ramani Department of Computer Sciences Institute for Cellular and Molecular Biology University of Texas at Austin

Biological Motivation • Human Genome Project has produced huge amounts of genetic data. • Next step is analyzing and interpreting this data.

Starting at the tip of chromosome 1... 1 taaccctaac cctaacccta accctaaccc taaccctaac cctaacccta accctaaccc 61 taaccctaac cctaacccta accctaaccc taaccctaac cctaacccaa ccctaaccct 121 aaccctaacc ctaaccctaa ccctaacccc taaccctaac cctaacccta accctaacct 181 aaccctaacc ctaaccctaa ccctaaccct aaccctaacc ctaaccctaa cccctaaccc 241 taaccctaaa ccctaaaccc taaccctaac cctaacccta accctaaccc caaccccaac 301 cccaacccca accccaaccc caaccctaac ccctaaccct aaccctaacc ctaccctaac 361 cctaacccta accctaaccc taaccctaac ccctaacccc taaccctaac cctaacccta 421 accctaaccc taaccctaac ccctaaccct aaccctaacc ctaaccctcg cggtaccctc 481 agccggcccg cccgcccggg tctgacctga ggagaactgt gctccgcctt cagagtacca 541 ccgaaatctg tgcagaggac aacgcagctc cgccctcgcg gtgctctccg ggtctgtgct 601 gaggagaacg caactccgcc ggcgcaggcg cagagaggcg cgccgcgccg gcgcaggcgc 661 agacacatgc tagcgcgtcg gggtggaggc gtggcgcagg cgcagagagg cgcgccgcgc 721 cggcgcaggc gcagagacac atgctaccgc gtccaggggt ggaggcgtgg cgcaggcgca 781 gagaggcgca ccgcgccggc gcaggcgcag agacacatgc tagcgcgtcc aggggtggag 841 gcgtggcgca ggcgcagaga cgcaagccta cgggcggggg ttgggggggc gtgtgttgca 901 ggagcaaagt cgcacggcgc cgggctgggg cggggggagg gtggcgccgt gcacgcgcag 961 aaactcacgt cacggtggcg cggcgcagag acgggtagaa cctcagtaat ccgaaaagcc 1021 gggatcgacc gccccttgct tgcagccggg cactacagga cccgcttgct cacggtgctg 1081 tgccagggcg ccccctgctg gcgactaggg caactgcagg gctctcttgc ttagagtggt ... 5641 gctccagggc ccgctcacct tgctcctgct ccttctgctg ctgcttctcc agctttcgct 5701 ccttcatgct gcgcagcttg gccttgccga tgcccccagc ttggcggatg gactctagca 5761 gagtggccag ccaccggagg ggtcaaccac ttccctggga gctccctgga ctggagccgg 5821 gaggtgggga acagggcaag gaggaaaggc tgctcaggca gggctgggga agcttactgt 5881 gtccaagagc ctgctgggag ggaagtcacc tcccctcaaa cgaggagccc tgcgctgggg 5941 aggccggacc tttggagact gtgtgtgggg gcctgggcac tgacttctgc aaccacctga 6001 gcgcgggcat cctgtgtgca gatactccct gcttcctctc tagcccccac cctgcagagc 6061 tggacccctg agctagccat gctctgacag tctcagttgc acacacgagc cagcagaggg 6121 gttttgtgcc acttctggat gctagggtta cactgggaga cacagcagtg aagctgaaat 6181 gaaaaatgtg ttgctgtagt ttgttattag accccttctt tccattggtt taattaggaa 6241 tggggaaccc agagcctcac ttgttcaggc tccctctgcc ctagaagtga gaagtccaga 6301 gctctacagt ttgaaaacca ctattttatg aaccaagtag aacaagatat ttgaaatgga 6361 aactattcaa aaaattgaga atttctgacc acttaacaaa cccacagaaa atccacccga 6421 gtgcactgag cacgccagaa atcaggtggc ctcaaagagc tgctcccacc tgaaggagac 6481 gcgctgctgc tgctgtcgtc ctgcctggcg ccttggccta caggggccgc ggttgagggt 6541 gggagtgggg gtgcactggc cagcacctca ggagctgggg gtggtggtgg gggcggtggg 6601 ggtggtgtta gtaccccatc ttgtaggtct gaaacacaaa gtgtggggtg tctagggaag... and 3x109 more...

Proteomics 101 • Genes code for proteins. • Proteins are the basic components of biological machinery. • Proteins accomplish their functions by interacting with other proteins. • Knowledge of protein interactions is fundamental to understanding gene function. • Chains of interactions compose large, complex gene networks.

Sample Gene Network

Yeast Gene Network Yeast ~5,800 genes ~5,800 proteins x 2-10 interactions/protein ~12,000 - 60,000 interactions ~10-20,000 known==>~1/3 of the way to a complete map!

Human Gene Network ~40,000 genes >>40,000 proteins x 2-10 interactions/protein <5,000 known ==> approx. 1% of the complete map! >>80,000 - 400,000 interactions ==> We’re a long ways from the complete map

Relevant Sources of Data Biological literature ~14 million documents DNA sequence data ~1010 nucleotides Gene expression data ~108 measurements, but... DNA polymorphisms ~107 known Gene inactivation (knockout) studies ~105 Protein structure data ~104 structures Protein interaction data ~104 interactions, but… Protein expression data ~104 measurements, but... Protein location data ~104 measurements

Extraction from Biomedical Literature • An ever increasing wealth of biological information is present in millions of published articles but retrieving it in structured form is difficult. • Much of this literature is available through the NIH -NLM’s Medline repository. • 11 million abstracts in electronic form are available through Medline. • Excellent source of information on protein interactions. • Need automated information extraction to easily locate and structure this information.

TI - Two potentially oncogenic cyclins, cyclin A and cyclin D1, share common properties of subunit configuration, tyrosine phosphorylation and physical association with the Rb protein AB - Originally identified as a ‘mitotic cyclin’, cyclin A exhibits properties of growth factor sensitivity, susceptibility to viral subversion and association with a tumor-suppressor protein, properties which are indicative of an S-phase-promoting factor (SPF) as well as a candidate proto-oncogene. Other recent studies have identified human cyclin D1 (PRAD1) as a putative G1 cyclin and candidate proto-oncogene. However, the specific enzymatic activities and, hence, the precise biochemical mechanisms through which cyclins function to govern cell cycle progression remain unresolved. In the present study we have investigated the coordinate interactions between these two potentially oncogenic cyclins, cyclin-dependent protein kinase subunits (cdks) and the Rb tumor-suppressor protein. The distribution of cyclin D isoforms was modulated by serum factors in primary fetal rat lung epithelial cells. Moreover, cyclin D1 was found to be phosphorylated on tyrosine residues in vivo and, like cyclin A, was readily phosphorylated by pp60c-src in vitro. In synchronized human osteosarcoma cells, cyclin D1 is induced in early G1 and becomes associated with p9Ckshs1, a Cdk-binding subunit. Immunoprecipitation experiments with human osteosarcoma cells and Ewing’s sarcoma cells demonstrated that cyclin D1 is associated with both p34cdc2 and p33cdk2, and that cyclin D1 immune complexes exhibit appreciable histone H1 kinase activity. Immobilized, recombinant cyclins A and D1 were found to associate with cellular proteins in complexes that contain the p105Rb protein. This study identifies several common aspects of cyclin biochemistry, including tyrosine phosphorylation and the potential to interact directly or indirectly with the Rb protein, that may ultimately relate membrane-mediated signaling events to the regulation of gene expression.

Manually Developed IE Systems for Medline • A number of projects have focused on the manual development of information extraction (IE) systems for biomedical literature. • KeX for extracting protein names (Fukuda et al., 1998):Extract words with special symbols excluding those with more than half of the characters being special symbols, hence eliminating strings such as “+/−”. • Suiseki for extracting protein interactions (Blaschke et al., 2001):PROT (0-2) PROT (0-2)complexNOUNbetween (0-3) PROT (0-3)and (0-3) PROT

Learning Information Extractors • Manually developing IE systems is tedious and time-consuming and they do not capture all possible formats and contexts for the desired information. • Machine learning from supervised corpora, is becoming the standard approach to building information extractors. • Recently, several learning approaches have been applied to Medline extraction (Craven & Kumlein, 1999; Tanabe & Wilbur, 2002; Raychaudhuri et al., 2002). • We have explored the use of a variety of machine learning techniques to develop IE systems for extracting human protein names and interactions, presenting uniform results on a single, reasonably large, human-annotated corpus.

Non-Learning Protein Extractors • Dictionary-based extraction • KEX (Fukuda et al., 1998)

Learning Methods for Protein Extraction • Rule-basedpattern induction • Rapier (Califf & Mooney, 1999) • BWI (Freitag & Kushmerick, 2000) • Token classification (chunking approach): • K-nearest neighbor • Transformation-Based Learning Abgene (Tanabe & Wilbur, 2002) • Support Vector Machine • Maximum entropy • Hidden Markov Models • Conditional Random Fields (Lafferty, McCallum, and Pereira, 2001) • Relational Markov Networks (Taskar, Abbeel, and Koller, 2002)

Our Biomedical Corpora • 750 abstracts that contain the word human were randomly chosen from Medline for testing protein name extraction. They contain a total of 5,206 protein references. • 200 abstracts previously known to contain protein interactions were obtained from the Database of Interacting Proteins. They contain 1,101 interactions and 4,141 protein names. • As negative examples for interaction extraction are rare, an extra set of30 abstracts containing sentences with non-interacting proteins are included. • The resulting 230 abstracts are used for testing protein interaction extraction.

The Yapex Corpus • 200 abstracts from Medline, manually tagged for protein names. • 147 randomly chosen such that they contain the Mesh terms “protein binding”, “interaction”, “molecular”. • 53 randomly chosen from the GENIA corpus http://www.sics.se/humle/projects/prothalt/

Evaluation Metrics for Information Extraction • Precision is the percentage of extracted items that are correct. • Recall is the percentage of correct items that are extracted. • Extracted protein names are considered correct if the same character sequences have been human-tagged as protein names in the exact positions. • Extracted protein interactions from an abstract are considered correct if both proteins have been human-tagged as interacting in that abstract. Positions are not taken into account.

Dictionary as Source of Domain Knowledge • Before applying machine learning, abstracts are tagged by matching n-grams against entries from a dictionary. Tagged abstracts are used as input for subsequent methods. • A dictionary of 42,000 protein names is used (synonyms included). • Generalization of protein names leads to increased coverage:

Rule-based Learning Algorithms: Rapier and BWI • Rule-based learning algorithms are used for inducing patterns for extracting protein names. • ForRapier (Califf & Mooney, 1999), each rule consists of a pre-filler pattern, a filler pattern and a post-filler pattern. [ human ] [ (2) transcriptase ] [ ( ] • For BWI (Freitag & Kushmerick, 2000), rules are composed of contextual patterns called wrappers, recognizing the start or end of a protein name. [ human ] [] [ transcriptase ] [ ( ] • High precision (> 70%) but low recall (< 25%).

NN:PROT … START END … START NN END NN Hidden Markov Models • We use part-of-speech information in HMMs as described in (Ray & Craven, 2001). • We train a positive model that generates sentences containing proteins, and a null model that generates sentences containing no proteins. • Select the model which gives the highest likelihood of generating a particular sentence, and tag the sentence using the Viterbi path in that model. • Moderate precision (~60%) and moderate recall (~40%).

Name Extraction by Token Classification(“Chunking” Approach) • Since in our data no protein names directly abut each other, we can reduce the extraction problem to classification of individual words as being part of a protein name or not. • Protein names are extracted by identifying the longest sequences of words classified as being part of a protein name. Two potentially oncogenic cyclins , cyclin A and cyclin D1 , share common properties of subunit configuration , tyrosine phosphorylation and physical association with the Rb protein

Constructing Feature Vectors for Classification • For each token, we take the following as features: • Current token • Last 2 tokens and next 2 tokens • Output of dictionary-based tagger for these 5 tokens • Suffix for each of the 5 tokens (last 1, 2, and 3 characters) • Class labels for last 2 tokens Two potentially oncogenic cyclins , cyclin A and cyclin D1 , share common properties of subunit configuration , tyrosine phosphorylation and physical association with the Rb protein

Maximum-Entropy Token Classifier • Distinguish among 5 types of tags: • S(-tart), C(-ontinue), E(-nd), U(-nique), O(-ther) • Feature templates: • current, previous, next word, and previous tag • part-of-speech for current, previous, next word • word class (full) ex: FGF1 => AAA0 • word class (brief) ex: FGF1 => A0 (Collins, ACL02) • An extraction’s confidence is the minimum of its transition probabilities. Example (4 tokens): t(y) is the forward probability of getting to state y at time step t

MaxEnt: Greedy Extraction • Use a Viterbi-like algorithm to find the most likely complete sequence of tags. • Drawback: many low confidence extractions are missed. • Want to be able to increase recall beyond Viterbi results to control precision-recall trade-off. • Solution: use a greedy extraction algorithm on all token sequences between any two consecutive Viterbi extractions.

Experimental Method • 10-fold cross-validation: Average results over 10 trials with different training and (independent) test data. • For methods which produce confidence in extractions, vary threshold for extraction in order to explore recall-precision trade-off. • Use standard methods from information-retrieval to generate a complete precision-recall curve. • Maximizing F-measure assumes a particular cost-benefit trade-off between incorrect and missed extractions.

Protein Name Extraction Results(Bunescu et al., 2004)

Graphical Models • An intuitive representation of conditional independence between domain variables. • Directed Models => well suited to represent temporal and causal relationships (Bayesian Networks, HMMs) • Undirected Models => appropriate for representing statistical correlation between variables (Markov Networks) • Generative Models => define a joint probability over observations and labels (HMMs) • Discriminative Models => specifies a probability over labels given a set of observations (Conditional Random Fields [Lafferty et al. 2001]). • Allow for arbitrary, overlapping features over the observation sequence.

Discriminative Markov Networks G = (V, E) – an undirected graph V= X  Y– a set of discrete random variables X – observed variables Y – hidden variables (labels) C(G) – the cliques of G Vc = Xc  Yc – the set of vertices in a clique cC(G) – the set of clique potentials A clique potential c specifies the compatibility of any possible assignment of values over the nodes in the associated clique c.

Conditional Random Fields [Lafferty et al. 2001] • CRF’s are a type of discriminative Markov networks used for tagging sequences. • CRF’s have shown superior or competitive performance in various tasks as: • Shallow Parsing • Entity Recognition • Table Extraction [Sha & Pereira 2003] [McCallum & Li 2003] [Pinto et al 2003]

Conditional Random Fields (CRFs)Lafferty, McCallum & Pereira 2001 • Undirected graphical model for sequence segmentation. • Log-linear model, different from MaxEnt model because of “global normalization” tags Start T1.tag T2.tag T3.tag Tn.tag End … tw … T1.w T2.w T3.w Tn.w cap T1.cap T2.cap T3.cap Tn.cap … • Tj.tag – the tag (one of S, C, E, U, O) at position j • Tj.w – true if word w occurs at position j • Tj.cap – true if word at position j begins with capital letter, …

Protein Name Extraction Results (Yapex)

Collective Classification of Web Pages [Taskar, Abbeel & Koller 2002]

Collective Information Extraction • Task: • Extracting protein/gene names from Medline abstracts. • Approach: • Collectively classify all candidate phrases from the same abstract. • Binary classification: • e.label = 0 => e is not a protein name • e.label = 1 => e is a protein name • Use two types of label correlations: • Acronyms and their long forms. • Repetitions of the same phrase.

Collective Information Extraction The control of human ribosomal protein L22 ( rpL22 ) to enter into the nucleolus and its ability to be assembled into the ribosome is regulated by its sequence . The nuclear import of rpL22 depends on a classical nuclear localization signal of four lysines at positions 13 – 16 … Once it reaches the nucleolus , the question of whether rpL22 is assembled into the ribosome depends upon the presence of the N - domain . e3 of rpL22depends repetition e5 e1 e2 overlap acronym repetition repetition L22 ribosomal protein L22 (rpL22) whetherrpL22is e4

e3 of rpL22depends e5 e1 e2 L22 ribosomal protein L22 (rpL22) whetherrpL22is e4 Relational Markov Networks [Taskar, Abbeel & Koller 2002] Discriminative Markov Networks, augmented with clique templates: • Overlap Template (OT) • Acronym Template (AT) • Repeat Template (RT)

Candidate Entities: Definition Candidate Entities: • The set of candidate entities usually depends on the type of named entity. • In general, could consider as candidates all phrases of length < L, where L may be task dependent. Two examples: • [Genes, Proteins] Most entity names are base noun phrases or parts of them. Thus a candidate extraction is any contiguous sequence of tokens whose POS tags are from {“JJ”, “VBN”, “VBG”, “POS”, “NN”, “NNS”, “NNP”, “NNPS”, “CD”, “”}, and whose head is either a noun or a number. • [People, Organizations, Locations] Most entity names are sequences of proper names potentially interspersed with definite articles and prepositions.

Candidate Entities: Local Features “… to the antioxidant superoxide dismutase  1 ( SOD1 ) enzyme and …” Entity Features: based on features introduced in [Collins ’02] • head word, with generic placeholder for numbers => “HD = 0” • entity text => “TXT = superoxide dismutase – 1” • entity type e.g. concatenation of its words types => “TYPE = a a – 0” • bigrams / trigrams at entity left / right boundaries based on combinations of lexical tokens, and word types. • Bigrams left => “BL = antioxidant superoxide”, “BL = antioxidant a”,… • Bigrams right => “BR = 0 (“,… • Trigrams left => “TL = the antioxidant superoxide”, “TL = the antioxidant a”,… • Trigrams right => “TR = 0 ( SOD1”, “TR = 0 ( A0”,… • suffix / prefix lists of words and word types • Preffixes => “PF = superoxide”, “PF = superoxide dismutase”,… • Suffixes => “SF = 0”, “SF = – 0”, “SF = dismutase – 0”,…

Overlap Template • Entity names should not overlap => hardwired overlap potential OT. e1 “… to the antioxidant superoxide dismutase  1 ( SOD1 ) enzyme and …” e2 e1 e2

u“eNOS” uOR0 uOR vOR v “eNOS” vOR v1 v2 v1 “eNOS interaction” v2 “eNOS interaction protein” … … v1 v2 vn u1 u2 um Repeat Template Production of nitric oxide ( NO ) in endothelial cells is regulated by direct interactions of endothelial nitric oxide synthase ( eNOS ) … Here we have used the yeast two - hybrid system and identified a novel 34 kDa protein , termed NOSIP ( eNOS interaction protein ) , which avidly binds to the carboxyl – terminal region of the eNOS oxygenase domain . u v

Acronym Template v2 v1 v d  “to the antioxidant superoxide dismutase  1 ( SOD1 ) enzyme and ” v3 vOR v1 v2 v3 v vOR … v1 v2 vn

Experimental Results • Datasets: • Yapex – a dataset of 200 Medline abstracts, manually tagged for protein names. • Aimed – a dataset of 225 Medline abstracts, of which 200 are known to mention protein interactions. • CoNLL – the CoNLL 2003 English dataset. Compared three approaches: • LT–RMN RMN extraction using local templates + Overlap Template • GLT–RMN  RMN extraction using both local and global templates. • CRF extraction as token classification using Conditional Random Fields[Lafferty et al 2001],with features based on current word, previous/next words, words short/long types and POS tags [Bunescu et al 2004].

Experimental Results – Yapex Yapex

Experimental Results – Aimed Aimed

Experimental Results – CoNLL CoNLL 2003

Protein Interaction Extraction • Most IE methods focus on extracting individual entities. • Protein interaction extraction requires extracting relations between entities. • Our current results on relation extraction have focused on rule-based learning approaches.

Rapier and BWI Revisited: the Inter-filler Approach • Existing rule-based learning algorithms are used for inducing patterns for identifying protein interactions. • Rules are learned for extracting inter-fillers.SHPTPWinteracts with another signaling protein,Grb7. • Inter-fillers are sometimes very long (~9 tokens on average; 215 tokens maximum!). For some rule-based learning algorithms (e.g. Rapier), the time complexity can grow exponentially in the length of inter-fillers.

Rapier and BWI Revisited: the Role-filler Approach • In the role-fillerapproach, we extract two interacting proteins into different slots, which we call the interactor and the interactee. • A sentence is divided into segments. Interactors are associated with interactees in the same segment using simple heuristics. We show that the S252W mutation allows the mesenchymal splice form of FGFR2 (FGFR2c) to bind and to be activated by the mesenchymally expressed ligands FGF7 or FGF10 and the epithelial splice form of FGFR2 (FGFR2b) to be activated by FGF2, FGF6, and FGF9. • Moderately high precision (> 60%) but low recall (< 40%).

Learning to Extract Proteins and their Interactions from Medline Abstracts