step by step
一. 文本预处理 Input: 文本摘要 Output: 得到文本摘要中含有两个或两个以上蛋白质的句子。 Procedure: 1. 对给的文本摘要进行分句。依据(.!?)进行划分句子。 2. 识别出每一个句子中的蛋白质。(这里需要改进, 需要借鉴CRF等Machine Learning等方法进行实体识别) 保存该句子、蛋白质名称、官方基因名称、以及在该句子中的位置、该句子中识别出蛋白质的个数 3. 将识别出的蛋白质转化为官方名称的基因名称。 将2中识别出来的蛋白质利用蛋白质-Gene转换成官方基因名称Gene 4. 输出包含至少两个Gene的句子。 For example:
->pombe Translin, which was cloned, revealed that it selectively binds d(GT)n and d(GTT)n microsatellite repeats.
->[pombe, Translin, ,, which, was, cloned, revealed, that, it, selectively, binds, d, (, GT, ), n, and, d, (, GTT, ), n, microsatellite, repeats, .]
->[pombe, TSN, ,, which, was, cloned, revealed, that, it, selectively, binds, d, (, ITGA2B, ), n, and, d, (, GTT, ), n, microsatellite, repeats, .]
->pombe TSN, which was cloned, revealed that it selectively binds d(ITGA2B)n and d(GTT)n microsatellite repeats.
二. 得到候选蛋白质相互作用对 使用Standford Parser对一中所得句子进行解析 -->> 得到语法树 使用Tregex定义规则结合关系词字典识别出如下三种情况下的相互作用关系.
Form 1: PIP, protein1 relation protein2.
Form 2: IPP, relation protein1 protein2.
Form 3: PPI, protein1 protein2 relation.
1. 识别出关系词;
2. 识别出否定词;
Three Rules: PIP, PPI, IPP, when there is a negative between the interation of PP
1. if PNIP (Protein .. Negative .. relation .. Protein), PINP, NPIP
P and P is not interaction. (Negative word is between the first Protein and interaction words)
for example:
1> proteinA is not interacted with ProteinB.
==>> PA I PB; ===>> not I ===>> PA I PB is false;
2> ProteinA and ProteinB is not interacted with ProteinC.
==>> PA I PC; PB I PC; ===>> not I ===>> PA I PC is false; PB I PC is false;
3> ProteinA is interacted with ProteinB, not with ProteinC.
==>> PA I PB; PA I PC; ===>> not PC ===>> PA I PC is false;
===>> PA I PB
4> ProteinA and ProteinB is interacted with ProteinC, not with ProteinD and ProteinE.
==>> PA I PC; PA I PD; PA I PE; PB I PC; PB I PD; PB I PE;
===>> not PD; not PE ===>> PA I PD is false; PA I PE is false; PB I PD is false; PB I PE is false;
===>> PA I PC; PB I PC;
2. if NPPI (Negative .. Protein .. relation .. Protein), PNPI, PPNI
not P and P interaction. (Negative word is ahead of the first protein)
for example:
1> not ProteinA and ProteinB interaction.
2> not ProteinA or ProteinB and ProteinC interaction.
3> not ProteinA is interacted with ProteinB.
4> not ProteinA is interacted with ProteinB and ProteinC.
3. if NIPP (Negative .. relation .. Protein .. Protein)
not interaction P and P. (Negative word is ahead of the relation word)
for example:
1> not interaction ProteinA and ProteinB.
4. the other situations:
1> Neither ProteinA nor ProteinB is interacted with ProteinC.
2> Neither ProteinA and ProteinB nor ProteinC with ProteinD is interacted with ProteinE.
3. 根据下面6项包括7条规则得到候选蛋白质的相互作用对;
1> IPP, PPI: 关系词I常为NN*(名词): 例如, A and B Interaction
PIP: 关系词I常为VP*(动词): 例如, A interacted with B
2> 三种形式中蛋白质实体之间词的个数.
PPI, PP之间常为1个词; IPP, PP之间常为1个或没有词. (这点不是很正确)
3> two proteins, a relation keyword; (similar with 2)
two proteins, a relation keyword, a negation keyword. For example:
a. A is not interacted with B.
b. A and B is not interaction.
c. not interaction A and B.
4> more than two proteins and a relation keyword.
use an algorithm: (Rule 5)
1) 标出句子中每个单词的位置, 从0开始.
2) 使用Hashtable保存蛋白质、关系词以及他们的位置.
3) 识别出hashtable中的关系词.
4) 通过结合关系词和出现在它前面和后面位置的蛋白质, 得到三元组.
5) 得到所有的PPIs.
5> more than two proteins, a relation keyword and a negation keyword.
类似于4, 只是需要额外检测 negation keyword.
例如: 蛋白质后跟着否定词可以被认为是false PPIs.
6> more than two proteins and two negation keywords.
一般这种情况是包含Neither/Nor的句子.
三. 蛋白质识别、蛋白质库的准确率
还有如下需要完成的: 二. 6> 、三
BUG: Pubmed id: 19740107 Description: 当句子长度很长时, 此时该如何处理? 直接截取前40个? or 将关系词进一步划分: PPI, IPP: 当关系词出现在两边时, 这两种情况下常出现的词有哪些? PIP: 中间的词经常是那些词? (这里的分词系统对词性分析不是很准确, 所以需要这样进一步划分)