Anyone interested in working on Graph LSTMs with me?

I’m interested in implementing a special LSTM architecture called “Graph LSTM” described in this paper : “Cross-Sentence N-ary Relation Extraction with Graph LSTMs” by Peng et al. with particular focus on their application of finding ternary relationships in multi-sentence texts.

They apply it to biomedical articles and that is a very interesting area for me. The goal is to extract the relationship between a Drug, Gene and Variant (Mutation) and determine if the patient would respond to the drug or not. I would want to expand the relationship to a QUARTERNARY relation, also including the disease or condition of the patient…

Example: “It is possible to conceive that an increase in the clinical doses of gefitinib or switching to erlotinib , which is given at its maximal tolerated dose [ ] , may lead to beneficial clinical effects , possibly by increasing BIM expression , in patients with EGFR mutations who acquired L747S after exposure to gefitinib . The data presented here indicate that BIM is both a marker and an effector of TKI induced apoptosis in EGFR-mutant NSCLC cells . Furthermore , we identified a novel acquired EGFR secondary mutation , L747S , and showed that both the L747S and the common T790M in cis to an activating EGFR mutation ( either L858R or an exon 19 deletion ) cause resistance to EGFR TKI induced apoptosis and attenuate the up-regulation of BIM.”

That sentence contains the following quaternary relationship

DRUGS: [afatinib || erlotinib)
MUTATIONS: [(L747S || T790M) && (L858R || exon 19 deletion)]
DISEASE: NSCLC (Non-Small-Cell Lung Carcinoma)

There is a nice description of the Graph LSTM in the paper, but it is a little too complicated for my Biologist brain to comprehend and know how to implement this model architecture in FASTAI.


(My intuition on this is that the idea of “Attention” is somewhat similar here, since rather than discovering a relationship between Sequence 1 and Sequence 2, there is a relationship between Sentence t-1 and sentence t (and Sentence t+1) in Sequence 1, but this may just be nothing…)

So, if anyone is interested in working on this with me, drop me a line!



What would be the advantages of using a Graph LSTM exactly. Just trying to understand it