A New Challenge for Text Mining: Cancer Risk Assessment

Motivation: Cancer Risk Assessment (RA) of chemicals is an important and challenging multi-step task which requires combining scientific expertise with elaborate literature search and review. Due to the rapidly growing volume of RA literature, the increasing complexity of experimental evidence, and the accelerating need for chemical assessment, the task is now getting increasingly challenging to manage via manual means. Text Mining (TM) technology specifically tailored for the needs of the task could lead to considerably more systematic and efficient RA. In this paper we present the first steps taken towards the development of such technology. Results: We have downloaded a corpus of 830 abstracts from PubMed and manually annotated the abstracts according to their relevance and the type of evidence they provide for cancer RA of selected test chemicals. The result is a taxonomy which classifies the key types of scientific evidence required for RA. The taxonomy can aid manual RA and is a starting point for the development of an approach based on TM. Using the annotated corpus we have demonstrated that supervised machine learning of large portions of the taxonomy and overall document relevance yields high accuracy and can be useful for the first step of cancer RA: finding the articles relevant for the task. We are now installing the automated classifier into the pipeline so that we can assess its impact on the RA process as a whole.