IBM Watson AI-enhanced search tool identifies novel candidate genes and provides insight into potential pathomechanisms of traumatic heterotopic ossification


Background: Traumatic heterotopic ossification (tHO) is the pathological formation of ectopic bone in soft tissues that can occur following injury to the skin, nervous system, or direct musculoskeletal trauma. Relatively high rates of tHO are expected after damage to neural structures. In clinical practice, diagnosis, prevention, and treatment of tHO are highly variable, partly due to a limited understanding of the pathophysiology. Identifying critical molecular contributors to the development of tHO remains challenging, limiting the development of effective diagnostics and treatment. IBM Watson for Drug Discovery (WDD) uses machine learning and natural language processing to interrogate a literature repository encompassing private and public data sources. This study used WDD to identify plausible new genes and pathways that may be involved in tHO.

Methods: A three-stage process centred around the disease agnostic WDD repository was applied during this study. Firstly, WDD was used to pool and target search the scientific literature involving heterotopic ossification arising from burns, orthopaedic trauma, and neurological injury populations. This training of the WDD natural language processing algorithms using known entities was used to discover novel intercepts in the network of semantic relationships evident in the published literature to 2019. Indications of plausible relationships were sought by triangulating biological concepts such as genes and diseases. In this step, using the WDD predictive analytics engine, the study identified and ranked 233 candidate genes that may be associated with pathological ectopic ossification, utilising a set of 100 genes with previously defined associations with tHO. Finally, a search of the WDD-linked literature related to the top 25 genes identified from the rank product analysis was conducted to validate WDD’s predictions of potential novel candidate genes.

Results: Of the top 25 ranked genes, six genes (MMRN1, MSC/MyoR, ITGAM/CD11b, PDGF-D, GREM-1 and NELL-1) were identified to have evidence of likely association with tHO. These candidate genes had previously defined roles in inflammation, aberrant tissue repair and regeneration, extracellular matrix remodelling and mineralisation, endochondral or intramembranous bone formation and injury-associated bone reactions, as well as functions in WNT and BMP signalling that are known to be important in osteogenic differentiation. Conclusions: Using a machine-learning approach, this study identified a novel set of plausible candidate gene targets associated with tHO. Machine-learning methods may effectively support target discovery and understanding of pathophysiology in complex disease states.

Link to Publisher Version (URL)


This document is currently not available here.

Find in your library