Extract relations between drugs and proteins (ReDL)

val documenter = new DocumentAssembler() 
    .setInputCol("text") 
    .setOutputCol("document")

val sentencer = new SentenceDetector()
    .setInputCols("document")
    .setOutputCol("sentences")

val tokenizer = new Tokenizer()
    .setInputCols("sentences")
    .setOutputCol("tokens")

val words_embedder = WordEmbeddingsModel()
    .pretrained("embeddings_clinical", "en", "clinical/models")
    .setInputCols(Array("sentences", "tokens"))
    .setOutputCol("embeddings")

val drugprot_ner_tagger = MedicalNerModel.pretrained("ner_drugprot_clinical", "en", "clinical/models")
    .setInputCols(Array("sentences", "tokens", "embeddings"))
    .setOutputCol("ner_tags") 

val ner_converter = new NerConverterInternal()
    .setInputCols(Array("sentences", "tokens", "ner_tags"))
    .setOutputCol("ner_chunks")

val pos_tagger = PerceptronModel()
    .pretrained("pos_clinical", "en", "clinical/models") 
    .setInputCols(Array("sentences", "tokens"))
    .setOutputCol("pos_tags")

val dependency_parser = DependencyParserModel()
    .pretrained("dependency_conllu", "en")
    .setInputCols(Array("sentences", "pos_tags", "tokens"))
    .setOutputCol("dependencies")

// Set a filter on pairs of named entities which will be treated as relation candidates
val drugprot_re_ner_chunk_filter = new RENerChunksFilter()
    .setInputCols(Array("ner_chunks", "dependencies"))
    .setMaxSyntacticDistance(10)
    .setOutputCol("re_ner_chunks")
    // .setRelationPairs(Array("CHEMICAL-GENE"))

// This model can also be trained on document-level relations - in which case, while predicting, use "document" instead of "sentence" as input.
val drugprot_re_Model = RelationExtractionDLModel()
    .pretrained("redl_drugprot_biobert", "en", "clinical/models")
    .setPredictionThreshold(0.9)
    .setInputCols(Array("re_ner_chunks", "sentences"))
    .setOutputCol("relations")

val pipeline = new Pipeline().setStages(Array(documenter, sentencer, tokenizer, words_embedder, drugprot_ner_tagger, ner_converter, pos_tagger, dependency_parser, drugprot_re_ner_chunk_filter, drugprot_re_Model))

val data = Seq("""Lipid specific activation of the murine P4-ATPase Atp8a1 (ATPase II). The asymmetric transbilayer distribution of phosphatidylserine (PS) in the mammalian plasma membrane and secretory vesicles is maintained, in part, by an ATP-dependent transporter. This aminophospholipid "flippase" selectively transports PS to the cytosolic leaflet of the bilayer and is sensitive to vanadate, Ca(2+), and modification by sulfhydryl reagents. Although the flippase has not been positively identified, a subfamily of P-type ATPases has been proposed to function as transporters of amphipaths, including PS and other phospholipids. A candidate PS flippase ATP8A1 (ATPase II), originally isolated from bovine secretory vesicles, is a member of this subfamily based on sequence homology to the founding member of the subfamily, the yeast protein Drs2, which has been linked to ribosomal assembly, the formation of Golgi-coated vesicles, and the maintenance of PS asymmetry. To determine if ATP8A1 has biochemical characteristics consistent with a PS flippase, a murine homologue of this enzyme was expressed in insect cells and purified. The purified Atp8a1 is inactive in detergent micelles or in micelles containing phosphatidylcholine, phosphatidic acid, or phosphatidylinositol, is minimally activated by phosphatidylglycerol or phosphatidylethanolamine (PE), and is maximally activated by PS. The selectivity for PS is dependent upon multiple elements of the lipid structure. Similar to the plasma membrane PS transporter, Atp8a1 is activated only by the naturally occurring sn-1,2-glycerol isomer of PS and not the sn-2,3-glycerol stereoisomer. Both flippase and Atp8a1 activities are insensitive to the stereochemistry of the serine headgroup. Most modifications of the PS headgroup structure decrease recognition by the plasma membrane PS flippase. Activation of Atp8a1 is also reduced by these modifications; phosphatidylserine-O-methyl ester, lysophosphatidylserine, glycerophosphoserine, and phosphoserine, which are not transported by the plasma membrane flippase, do not activate Atp8a1. Weakly translocated lipids (PE, phosphatidylhydroxypropionate, and phosphatidylhomoserine) are also weak Atp8a1 activators. However, N-methyl-phosphatidylserine, which is transported by the plasma membrane flippase at a rate equivalent to PS, is incapable of activating Atp8a1 activity. These results indicate that the ATPase activity of the secretory granule Atp8a1 is activated by phospholipids binding to a specific site whose properties (PS selectivity, dependence upon glycerol but not serine, stereochemistry, and vanadate sensitivity) are similar to, but distinct from, the properties of the substrate binding site of the plasma membrane flippase.""").toDS.toDF("text")

val result = pipeline.fit(data).transform(data)