Drug treatment extraction (Deliverable 1)

- Build first-pass lexicon of "treatment" synonyms

- Refine lexicon of "treatment" synonyms (use pre-trained Word2Vec model?)

- Get training data for drug-"treats" indication task

- Extract drug-"treats" indication in COVID papers with co-occurrences

- Extract drug-"treats" indication in COVID papers with syntax (dependency) parsing

- Extract drug-"treats" indication in COVID papers with DL (approach 1 - BERT)

- Extract drug-"treats" indication in COVID papers with DL (approach 2)

- Evaluate performance of all drug-"treats" models, select best model

- Incorporate best drug-"treats" model into final pipeline

- Hedge detection in treatment sentences using co-occurrence

- Hedge detection in treatment sentences using syntax parsing

- Hedge detection in treatment sentences using DL

- Evaluate performance of all hedge detection models, select best model

- Incorporate best hedge detection model into final pipeline

- Negation detection in treatment sentences using co-occurrence

- Negation detection in treatment sentences using syntax parsing

- Negation detection in treatment sentences using DL

- Evaluate performance of all negation detection models, select best model

- Incorporate best negation detection model into final pipeline

- Refine the drug lexicon (use intersection of multiple lexicons, pre-filter somehow?)

- Collect metadata about drugs

- Join and incorporate metadata into final pipeline

Model organism extraction (Deliverable 1)

- Explore animal model mentions in CORD-19 (will spawn subtasks)

Paper-level annotation of "computational", "experimental", "clinical" (Deliverable 1)

- Manually annotate 1000 papers as containing "computational", "experimental (lab)", or "clinical" evidence

- Automatically annotate papers' evidence types based on co-occurrence

- Automatically annotate papers' evidence types using an unsupervised learning approach

- Automatically annotate papers' evidence types based on DL (approach 1 - BERT)

- Automatically annotate papers' evidence types based on DL (approach 2)

- Evaluate performance of all evidence classifiers, select best model

- Incorporate best evidence classifier model into final pipeline

- Extract methods being used using topic modeling over "Methods" section of paper (TF-IDF?)

- Contingent on success of 1T.3.5.1, incorporate into final pipeline

COVID mechanism of action knowledge graph creation (Deliverable 2)

- Create a Neo4J KG from Kroger paper*

- Consider incorporating STRING PPI relations into Neo4J

- Get training data for relation extraction from text

- Extract PPI relations from text using co-occurrence

- Extract PPI relations from text using syntax parsing

- Extract PPI relations from text using ML (approach 1)

- Extract PPI relations from text using BERT (approach 2)

- Explore tissue specificity of PPI relations in BERT (will spawn subtasks)

- Explore drug-protein interactions in CORD dataset (will spawn subtasks)

- Incorporate all relationships extracted into Neo4J KG

*Kroger paper: https://www.biorxiv.org/content/10.1101/2020.03.22.002386v1.full.pdf

Posology information extraction (Deliverable 3)

- Filter documents (primary clinical literature)

- Extract Dosing Information (MedEx)

- Filter MedEx annotation errors (get rid of irrelevant info)

- Extract clinical trial ID numbers

- Evaluate results (for each step)

- Repeat pipeline SST 1T.5.1-1T.5.5 for all posology columns (will possibly spawn subtasks)

Task V&T infrastructure

- Increase formality of language in testing framework doc on GitHub wiki

- Implement improved testing framework

- Brainstorm and set up an evaluation pipeline for ML components

- Clean up GitHub repo (strip functions into helper functions)