CN104657626A - Method for constructing protein interaction network by using text data - Google Patents

Abstract

The invention discloses a method for constructing a protein interaction network by using text data, which is characterized by comprising the following steps: establishing a protein set; secondly, recording the probability value of interaction of every two proteins in the protein set; thirdly, constructing an initial network structure according to the size of the probability value; sixthly, repeatedly selecting protein, giving a positive or negative action feedback value, and continuously iterating on the initial network structure to obtain a final network structure. The invention adopts a mode of repeated selection and interaction, constructs a probability graph of an action network through reinforcement learning based on positive feedback, negative feedback and feedback inhibition, and seamlessly combines the probability graph with biological knowledge and biological data.

Description

A kind of method utilizing text data to build protein-protein interaction network

Technical field

The present invention relates to a kind of field of biology, particularly relate to a kind of method utilizing text data to build protein-protein interaction network.

Background technology

Biosystem comprises the network of a lot of different aspects and different tissues form.The most important feature of life system complicacy is not only the complicacy of its constituent, is more the complicacy of relation between each constituent.So, when analysing biomolecules network, not only need each molecular entity in abundant awareness network, the more important thing is the mutual relationship understood between each molecular entity.Protein is the important biomolecule of a class, forming by interaction each other the links that protein-protein interaction network participates in bio signal transmission, Gene expression and regulation, energy and the life process such as metabolism and cell cycle regulating, is the basis that a lot of biological function realizes.Interaction between protein is forming nearly all life system, is regulating and controlling in various physiology/pathogenesis to play vital effect.The biological function that protein interaction is not only research agnoprotein matter provides clue, is also the biological mechanism fully understanding a cell or a biological approach, provides necessary information.In biomedicine, Study on Protein interphase interaction has very important realistic meaning.The interaction relationship of the protein of systematic analysis class disease association, for the principle of work understanding these protein in biosystem, understand the reaction mechanism of bio signal and metabolism of energy substances under special physiological state, and the functional cohesion understood between disease-associated protein is all significant.

At present, there is multiple method for building protein-protein interaction network, mainly be included in high flux experiment basis on set up interactive network, utilize existing data mining interactive network in document, the method establishment interactive network etc. predicted by computing technique.But, in general, a lot of method Shortcomings building protein-protein interaction network.

First, the restriction that protein-protein interaction network generally can be subject to expense is set up on the basis of high flux experiment.The method of a lot of high flux experiment is when studying certain disease, remain and be confined to a small amount of protein, not do not remove structure from the angle of protein profiling widely and analyze, its main cause is, the biochemical test of analysing protein interphase interaction costly, result in and can only choose a small amount of protein, cannot analyze and research using all protein as candidate albumen matter widely.And choose a small amount of protein and analyze and research, not only very likely omit the protein with this disease association, miss that some are biomedical true, and the visual angle of analysis and research and thinking can be limited to, more difficult discovery fresh information and new knowledge.

Secondly, the method establishment protein-protein interaction network of data in literature is utilized merely can be subject to the impact of the quality of data and associated biomolecule analysis thereof.Sometimes the data deriving from different document can make different biology explanations and conclusion to same biological phenomenon; And same batch data has different biology explanations and conclusion sometimes.This goes to analyze same phenomenon from different perspectives can produce different explanations and conclusion because people understand to result in complex biological phenomenon not comprehensively.Therefore complex biological problem is being researched and analysed, during as built protein-protein interaction network, needing data and the relevant information of fully integrating separate sources, various information is screened, eliminate the false and retain the true, thus deepen the complete understanding in and profound level multi-level to its pathogenic mechanism.

In addition, a lot of computing method building protein-protein interaction network bias toward design and the improvement of computation model, but fail to merge biological knowledge and the biology fact well, to such an extent as to occur the wrong conclusion that some and biological ABC and the fact are runed counter to.

Summary of the invention

The object of the invention is to provide a kind of method utilizing text data to build protein-protein interaction network, existing biological field knowledge can be merged, the data that rear era gene obtains can be made full use of again, take into account the new protein-protein interaction network establishing method of complex network characteristic simultaneously.

For achieving the above object, the technical solution used in the present invention is: a kind of method utilizing text data to build protein-protein interaction network, comprising:

(1) set up protein set;

(2) record all proteins in protein set and interactional probable value occurs between two;

(3) build initial network structure according to the size of probable value;

(5) repeatedly select protein, given positive interaction or negative interaction value of feedback, in initial network structure, continuous iteration, obtains the network structure of final protein-protein interaction network.

Technique scheme is, described " interactional probable value occurs all proteins between two " is, in protein set, select arbitrarily a protein as main mutual protein, be by mutual protein with other protein, the mutual protein of described master and each is mutual by mutual protein, form an interactively, the then more mutual protein of change owner, again undertaken alternately with other by mutual protein, form another interactively, circulation like this, cycle index reaches predetermined value, and when repeating to select, calculate in an iterative manner, obtain final interactively as the mutual probable value of corresponding two protein.

Technique scheme is, described, and " repeating situation about selecting " is, the some protein in protein set and another protein are mutually as leading, by the interactive situation of mutual protein, and repeat mutually to be selected interactive situation again.

Further technical scheme, described predetermined value is: the protein in each set was all undertaken mutual as main mutual protein and other by mutual protein, or no longer include renewal in the long period section that circulates, or reach one or more in specified iterative steps.

Technique scheme is, described structure initial network structure is: each protein in protein set is as a node, and occur between two to interact as limit, its boundary values is larger, then there is interactional probability between any two larger, otherwise then less, in the process built, large being enhanced alternately of boundary values, until no longer include renewal in a long period section, otherwise be then weakened, until probable value is zero, finally obtain the network structure built by node and limit; By the initial network structure built, then obtain final network structure further.

Further technical scheme, described final network structure is: by using entropy assessment to build network, calculate the entropy weight of each protein node, then computational grid entropy weight, and entropy weight is less, represents network stabilization, upgrades initial network structure.

Technique scheme is, the foundation of described protein set:

A. by obtaining required text in biomedical literature database;

B. mutually do to obtain protein name and identification number thereof relational database from protein;

C. according to the protein name that step b obtains, identify the protein name in the described text that obtains in step a, and mark corresponding identification number;

D. described protein set P={p is built _i, wherein p _irepresent the identification number corresponding to i-th protein.

Because technique scheme is used, the present invention compared with prior art has following advantages:

1. the present invention adopts and to make repeated attempts reciprocation, increases or weaken the boundary values of pairwise interaction, and the network structure of structure occurs as the result of dynamic, ensure that the uncalibrated visual servo characteristic that complex biological network should possess;

2. adopt construction method of the present invention, meet the feature of the non-intellectual of biological questions, in unknown random environment, obtain best behavior, build and there is the protein-protein interaction network of non-intellectual, can ensure that network convergence is to a best steady state (SS);

3. be seamlessly combined with biological knowledge and biological data in the process setting up network, enhanced biological is true, and nonrandom structure network, guarantee the fundamental characteristics of network conforms biological complex network.

Accompanying drawing explanation

Fig. 1 utilizes text data to build protein-protein interaction network method implementation step process flow diagram;

Fig. 2 utilizes text data to use the intensified learning method of Average Reward value to build the node degree probability distribution schematic diagram of protein-protein interaction network;

Fig. 3 utilizes text data to use the intensified learning method of Average Reward value to build the node degree probability density distribution schematic diagram of protein-protein interaction network.

Embodiment

Below in conjunction with drawings and Examples, the invention will be further described:

Embodiment one: shown in Figure 1, a kind of method utilizing text data to build protein-protein interaction network, comprising:

(1) set up protein set;

(2) record all proteins in protein set and interactional probable value occurs between two;

(3) build initial network structure according to the size of probable value;

(4) repeatedly select protein, given positive interaction or negative interaction value of feedback, in initial network structure, continuous iteration, obtains the network structure of final protein-protein interaction network.

The foundation of described protein set:

A. by obtaining required text in biomedical literature database;

B. mutually do to obtain protein name and identification number thereof relational database from protein;

C. according to the protein name that step b obtains, identify the protein name in the described text that obtains in step a, and mark corresponding identification number;

D. described protein set P={p is built _i, wherein p _irepresent the identification number corresponding to i-th protein.

Described " interactional probable value occurs all proteins between two " is, in protein set, select arbitrarily a protein as main mutual protein, be by mutual protein with other protein, the mutual protein of described master and each is mutual by mutual protein, form an interactively, the then more mutual protein of change owner, again undertaken alternately with other by mutual protein, form another interactively, circulation like this, cycle index reaches predetermined value, and when repeating to select, calculate in an iterative manner, obtain final interactively as the mutual probable value of corresponding two protein.

Described " repeat select situation " is, the some protein in protein set and another protein are mutually as leading, by the interactive situation of mutual protein, and repeat mutually to be selected interactive situation again.

Described predetermined value is: the protein in each set was all undertaken alternately by mutual protein as main mutual protein and other, or no longer includes renewal in the long period section that circulates, or reaches one or more in specified iterative steps.

Described structure network structure mode is: each protein in protein set is as a node, occur between two to interact as limit, its boundary values is larger, then there is interactional probability between any two larger, otherwise it is then less, in the process built, large being enhanced alternately of boundary values, until no longer include renewal in a long period section, otherwise be then weakened, until probable value is zero, finally obtain the network structure built by node and limit, this network structure is effect network is occur as the result of learning behavior dynamic.

Described final network structure is: by using entropy assessment to build network, calculate the entropy weight of each protein node, then computational grid entropy weight, and entropy weight is less, represents network stabilization, upgrades initial network structure.

What adopt in the present embodiment is that intensified learning method builds network structure, and in the framework of intensified learning, set up protein-protein interaction network, node represents protein, is designated as node 1 ..., node n, while the effect representing between protein.A node, under the decision-making of intensified learning agent, obtains an action, and this action may be that this protein and other protein exist cooperative relationship, has interaction between two protein that expression is relevant; Also may be that this protein and other protein exist mutex relation, between two protein that expression is relevant, can not interaction be had; Also likely can not determine whether have interaction between two relevant protein.Node all can obtain an award at every turn after carrying out mutual trial, and the value of award determines which will be enhanced alternately.Repeatedly select.Along with the propelling of time, protein adjustable strategies, also can decision policy again, introduces randomness simultaneously, explores, to conform.When obtaining the result of unsatisfactory result, can change strategy be selected, or select other protein of change.Like this, both allow for the evolution of protein-protein interaction network, contemplate the evolution of individual strategy.Final protein-protein interaction network occurs as the result of agent learning behavior dynamic.

Certain node i Stochastic choice accesses other nodes, select probability by each node by other nodes the relative right to choose re-computation of giving obtain.Each node selects the strategy of accessing other nodes very well, and has a strengthening at every turn.Node i has one to select weight vectors < w _i1 ..., w _in > calculates the probability selecting other each points, and its mode calculated is the new node added due to each time connects already present node i with certain probability, so any one node i is at the selection weight vectors w of t _it () is a stochastic variable, and if node i be w in the selection weight in t-1 moment _i(t-1), then it at the selection weight w of t _it () only depends on its selection weight in the t-1 moment.Can new node be connected when t carves, have nothing to do with the history before the t-1 moment.

Most nitrification enhancement has one to estimate the function of this state (or performing given action in state) fine or not degree when a given state (or state action to) to agent, is called value function.Function V ^hfor the state value function of tactful h.Under tactful h, take the value of action u for from state x at state x, take action u, then follow the expected returns of tactful h, be designated as Q ^h(x, u).Q ^πfor the working value function of tactful h, be used for weighing the fine or not degree taking action u at state x.

V value and Q value need to upgrade along with time step.Conventional method utilizes the accoumulation of discount to award.But the some shortcomings that this method exists, manually determine the setting of discount factor, parameter and relevant to application etc. as needed.In the process of network evolution, the sequencing that the formation on the limit between the node of expression interactively should occur with it has nothing to do.But, in practical situations both, there is several factors evolutional sequence can be caused different, as the sequencing etc. that data are read in.But in the structure of interactive network, no matter how middle evolutional sequence changes, and same group of data, identical method, should obtain finally consistent result.Therefore, the method using accoumulation of discount award is not suitable for.Given this, need to use a kind of Q value of haveing nothing to do with network evolution order and V value calculating method to assess constructed network.

We weigh randomness or scrambling, to measure the stability of network by using entropy.Entropy is larger, and randomness is larger.And entropy is less, then randomness is less, meets the variable condition of biosystem.If wd _iwhat represent node i adds measures and weights (weighted degree, wd), then local entropy (localentropy, the le) definition of node i as shown in Equation 1.

$\mathrm{le}\left(i\right)=\frac{1}{\log {\mathrm{wd}}_i}\underset{j\&Element;N\left(i\right)}{\mathrm{\&Sigma;}}{w}_{\mathrm{ij}}\log w_{\mathrm{ij}}---\left(1\right)$

Wherein, wd _ithe weight sum of having an effect of all nodes relevant to node i, w _ijit is the weight on the limit between node i and node j.

A network of network entropy (network entropy, ne) is the entropy sum of all nodes, as shown in Equation 2.

$\mathrm{ne}=\underset{i\&Element;V}{\mathrm{\&Sigma;}}\mathrm{le}\left(i\right)---\left(2\right)$

Through long-term iteration, the protein-protein interaction network of final formation is not random network, is to have a stable topological structure, therefore, the minimum network entropy that optimum topological protein interaction has, thus final network structure the most stable can be obtained.

Specific implementation step is:

Step (1): the E-utility interface using biomedical literature database PubMed to provide obtains required text from biomedical literature database PubMed;

Step (2): download from protein interaction relationship data database DIP, IntAct and STRING and obtain protein name and identification number thereof;

Step (3): identify the protein name in text, uses identification number to represent;

Step (4): user provides the protein set P={p in the protein-protein interaction network needing to build _i, wherein p _irepresent the identification number corresponding to i-th protein;

Step (5): get protein set P={p _iin all any two protein, form candidate protein effect to collection all_pairs;

Step (6): the protein effect of setting available candidate is to collection avaiable_pairs=all_pairs;

Step (7): if the protein active set avaiable_pairs of available candidate also has untreated effect right, appoints and gets one of them effect to (p _i, p _j), enter next step, otherwise proceed to step (14);

Step (8): from the protein effect of available candidate to concentrated removal effect to (p _i, p _j), avaiable_pairs=avaiable_pairs-{ (p _i, p _j);

Step (9): initialization protein p _iwith protein p _jthere is interactional weight weight (p _i, p _j)=0.0;

Step (10): search for protein p respectively in protein interaction relationship data database DIP, IntAct and STRING _i,p _jbetween interaction situation;

Step (11): if having protein p in DIP database _i, p _jbetween interaction, then weight (p _i, p _j)=weight (p _i, p _jthe reward value of)+presetting; Otherwise, if clearly represent protein p in DIP database _i, p _jbetween not do not interact, then weight (p _i, p _j)=weight (p _i, p _jthe penalty value of)-presetting; Protein p is not searched else if in DIP database _i, p _jthere is interactional information, then weight (p _i, p _j) value remains unchanged;

Step (12): if having protein p in IntAct database _i, p _jbetween interaction, then weight (p _i, p _j)=weight (p _i, p _jthe reward value of)+presetting; Otherwise, if clearly represent protein p in IntAct database _i, p _jbetween not do not interact, then weight (p _i, p _j)=weight (p _i, p _jthe penalty value of)-presetting; In IntAct database, not searching protein pi else if, there is interactional information in pj, then weight (p _i, p _j) value remains unchanged;

Step (13): if having protein p in STRING database _i, p _jbetween interaction, then weight (p _i, p _j)=weight (p _i, p _jthe reward value of)+presetting; Otherwise, if clearly represent protein p in STRING database _i, p _jbetween not do not interact, then weight (p _i, p _j)=weight (p _i, p _jthe penalty value of)-presetting; Protein p is not searched else if in STRING database _i, p _jthere is interactional information, then weight (p _i, p _j) value remains unchanged;

Owing to containing abundant biological field knowledge in protein interaction relationship data database DIP, IntAct and STRING, by the setting of initial value, the weight of the protein interaction of Given information can be heightened, the known effect that protein interaction can not occur is reduced weight.

Step (14): obtain the protein effect network and the initializes weights matrix that are rich in biomedical knowledge, N=(p _i, p _j, weight (p _i, p _j)) in;

Step (15): initialization candidate albumen matter collection candidate_protein, adds initialization candidate albumen matter collection by all proteins;

Step (16): an optional protein p from candidate albumen matter collection candidate_protein _i;

Step (17): remove protein p from candidate albumen matter collection candidate_protein _i, candidate_protein=candidate_protein-{p _i;

Step (18): be initialized to protein collection candidate_pair_protein, adds initialization candidate albumen matter collection by all proteins;

Step (19): if protein collection candidate_pair_protein is not empty in pairs, then an optional paired protein p from candidate's paired protein collection candidate_pair_protein _j; Otherwise forward step (17) to;

Step (20): utilize formula calculate current network entropy;

Step (21): use Greedy strategy to select protein p _i, p _jbetween whether have interaction;

Step (22): if Q _fbe less than Q ', then think p _i, p _jbetween not do not interact, weight (p is set _i, p _j)=0.0; Otherwise think p _i, p _jbetween have interaction, weight (p _i, p _j)=Q _f;

Step (23): more novel protein p _i, p _jbetween have interactional probability to be

Step (24): use new weight (p _i, p _j) value, upgrade protein-protein interaction network N;

Step (25): after reaching specified iterative steps, no longer upgrade, obtain final network structure.

End condition can be that matrix weight does not upgrade or reached predetermined iterative steps in a long period section.Matrix weight may be used for the selection of action, the i.e. selection of node interphase interaction, its select probability is: the matrix weight therefore finally obtained can be considered as topology of networks, and the renewal process of matrix weight can regard the evolutionary process building network as.

Claims (7)

1. utilize text data to build a method for protein-protein interaction network, it is characterized in that, comprising:

(1) set up protein set;

(2) record all proteins in protein set and interactional probable value occurs between two;

(3) build initial network structure according to the size of probable value;

(4) repeatedly select protein, given positive interaction or negative interaction value of feedback, in initial network structure, continuous iteration, obtains the network structure of final protein-protein interaction network.

2. protein-protein interaction network construction method according to claim 1, it is characterized in that: described " interactional probable value occurs all proteins between two " is, in protein set, select arbitrarily a protein as main mutual protein, be by mutual protein with other protein, the mutual protein of described master and each is mutual by mutual protein, form an interactively, the then more mutual protein of change owner, again undertaken alternately with other by mutual protein, form another interactively, circulation like this, cycle index reaches predetermined value, and when repeating to select, calculate in an iterative manner, obtain final interactively as the mutual probable value of corresponding two protein.

3. protein-protein interaction network construction method according to claim 2, it is characterized in that: described " repeating situation about selecting " is, some protein in protein set and another protein mutually as main, by the interactive situation of mutual protein, and repeat mutually to be selected interactive situation again.

4. protein-protein interaction network construction method according to claim 2, it is characterized in that: described predetermined value is: the protein in each set was all undertaken mutual as main mutual protein and other by mutual protein, or no longer include renewal in the long period section that circulates, or reach one or more in specified iterative steps.

5. protein-protein interaction network construction method according to claim 1, it is characterized in that: described structure initial network structure is: each protein in protein set is as a node, occur between two to interact as limit, its boundary values is larger, then there is interactional probability between any two larger, otherwise it is then less, in the process built, large being enhanced alternately of boundary values, until no longer include renewal in a long period section, otherwise be then weakened, until probable value is zero, finally obtain the network structure built by node and limit; By the initial network structure built, then obtain final network structure further.

6. protein-protein interaction network construction method according to claim 5, it is characterized in that: described final network structure is: build network by using entropy assessment, calculate the entropy weight of each protein node, computational grid entropy weight again, entropy weight is less, represent network stabilization, upgrade initial network structure.

7. protein-protein interaction network construction method according to claim 1, is characterized in that: the foundation of described protein set:

A. by obtaining required text in biomedical literature database;

B. mutually do to obtain protein name and identification number thereof relational database from protein;

C. according to the protein name that step b obtains, identify the protein name in the described text that obtains in step a, and mark corresponding identification number;

D. described protein set P={p is built _i, wherein p _irepresent the identification number corresponding to i-th protein.