ASKE-E Month 11 Milestone Report

Integration with ASKE modeling frameworks

We collaborated with other teams to decide on a unified modeling framework to simulate and visualize different models built in ASKE the same way. As a result we designed a process to convert PySB reaction networks of EMMAA models into the PetriNet Classic GroMEt format developed by University of Arizona team. The GroMEt structure includes State and Rate junctions connected by wires. In the context of EMMAA models, State and Rate junctions are represented by model species and reaction rates respectively. Wires include the connections from reactants to the rates and from rates to products. After discussions with the University of Arizona and Uncharted teams on what metadata is necessary for meaningful visualizations, we added custom metadata to GroMEts generated from EMMAA models that includes mappings from State junctions to INDRA Agents and from Rate junctions to INDRA Statements and PySB rules.

Generation of GroMEts is now deployed as a part of automated update pipeline and their daily updated JSON exports are available for download on S3 and on the EMMAA dashboard. We also uploaded GroMEt exports for two different EMMAA models (the MARM model and the Ras Machine 2.0 model) to the shared GitHub repo maintained by the Galois team for the upcoming ASKE-E final demo.

BioCreative participation

The BioCreative challenge is a longstanding community effort to evaluate text mining systems applied to biology. This year, BioCreative includes a special track for COVID-19 text mining tool interactive demos which focuses on text mining-based tools specifically developed to support COVID-19 research efforts. We registered for this track with a proposal on the EMMAA COVID-19 model titled “A self-updating causal model of COVID-19 mechanisms built from the scientific literature”, and our proposal was accepted for participation. We also received some useful feedback on how to improve the EMMAA model query interface and the statement browser interface which we subsequently implemented (as described in this report). Going forward, we will continue to improve the EMMAA COVID-19 model and surrounding features, and aim to highlight ways in which EMMAA goes significantly beyond just text mining and knowledge assembly, encompassing also automated modeling and data analysis based on text mining results.

Improving the EMMAA model query interface

In the previous report we shared the updates on the addition of new query types and improvements in the interactive query interface. This month we extended the tutorial on using the query UI. We added sections about navigating different parts of query page and selecting the correct query type based on the scientific question and updated the descriptions and examples for all supported types of queries.


Part of updated query tutorial

We exposed the links to both written tutorials and video demonstations of the tool on the query page.


Links to demos and tutorials from query page

Improving the EMMAA statement browser

We extended the set of features for browsing all statements in a given EMMAA model. It can be often useful to focus on one type of interaction when browsing or curating statements. To enable this, we added a filter by statement type that is shown in the image below.


COVID-19 model statements filtered to Inhibition

In addition to filtering statements by type from the all statements view, users can also click on any of the horizontal bars on the statement types distribution plot on the EMMAA model dashboard to be redirected to a page displaying statements filtered down to that type.


Statement types distribution chart before clicking to open statements view

Previously we supported sorting the statements by the number of unique evidences they have and by the number of paths they support. Recently we also added an option to sort statements by their belief score.

Using custom belief scorers for EMMAA models

During this period we have developed an approach to deploying custom probability models to estimate the reliability (“belief”) of statements in EMMAA models. As part of our ongoing efforts to validate and improve the accuracy of these belief estimates, we have developed and validated several machine learning models (e.g. logistic regression, random forest) to empirically estimate belief based on a dataset of roughly 5,000 statements that we have manually curated. A valuable feature of these models is that they can capture the role of features other than reader evidence counts in estimating belief; for example, we have found that statement type and number of unique supporting PMIDs are also informative. We have also extended this approach to include “hybrid” models that incorporate machine learning for estimating reliability from text mining sources and a set of priors for curated databases.

We created a framework for deploying versioned, alternative belief models to S3 after training and subsequently making use of them during the assembly of EMMAA models. The EMMAA model configuration now takes a user-configurable parameter specifying which belief model to load and use. Statement belief estimates are now also displayed in the front-end and can be used to sort statements in the All Statements view (see screenshot below).


All statements view for the BRCA model, showing the orange belief badges on the right

We are working to draw on additional statement evidences that are in the INDRA Database (but outside the scope of the EMMAA model) to enhance estimates of belief. This way, a statement that may appear rarely in text for a specific disease context can be corroborated by information appearing outside that context, such as in a pathway database or in papers not incorporated by the EMMAA model. This will separate the technical estimate of a statement’s reliability from its canonicalness in a specific context, allowing users to identify high-confidence extractions that may be novel in the context of a particular disease.

To demonstrate these new developments, we computed belief estimates for the neurofibromatosis model in four different configurations: with the default belief model vs. a new, partly machine-learned “hybrid” model, and with EMMAA-only evidences vs. evidences from both EMMAA and the INDRA DB. As shown in the figure below, the inclusion of additional evidence from the INDRA DB shifts belief estimates to the right due to the addition of extra evidence, while the hybrid model provides a more continuous stratification of belief than the default belief model. In the upcoming period we will evaluate the use of this approach in other models and determine whether the new belief estimates are well-calibrated.


Belief scores of statements in the EMMAA model, using the default belief model (left plots) or random-forest-based hybrid model (right plots); and using only EMMAA evidence (top plots) or including evidences from the INDRA DB (bottom plots).

Developments in relation extraction from text

We have previously reported on completing our goals to enable named entity recognition and grounding in the Reach reading system for (1) viral proteins (2) human and non-human (including viral) protein chains and fragments, and have developed new algorithms in INDRA for organism disambiguation for proteins in the context of a given publication.

This month, we continued our work on creating a training data set for recognizing causal precedence in text. The goal is to find a set of positive and negative examples where a paper describes an A-B interaction and also a B-C interaction, and an A->B->C causal chain is implied (in the positive case) or not implied (in the negative case). This labeled data can then be used to train a classifier that can be run on elementary relation extractions to reconstruct causal precedence relations. We have previously reported on our approach to automatically finding positive examples. Since then, we have worked on an alternative approach to finding negative examples. First, we searched for papers in which both the A-B and the B-C relationship could be found within a specified distance of each other. To improve the quality/reliability of each example, we also implemented a filter to retain only A-B, B-C pairs where each is supported by additional background evidence beyond the given paper (this helps eliminate text mining errors). We then reviewed the results to curate positive vs negative examples. We found that the vast majority of examples remaining were positive for causal precedence. This imples that proximity in text may often be sufficient to infer causal precedence across A-B, B-C relations. We are investigating this further while continuing to develop an improved method for finding negative examples.