Neogenomics, Inc. · NASDAQ:NEO
Neogenomics — Technical Presentation (2020)
10 slides from Neogenomics's technical presentation, 2020
Slide 1: A highly sensitive and specific gene fusion algorithm using ensemble machine learning
A highly sensitive and specific gene fusion algorithm based on multiple fusion callers and an ensemble machine learning approach K Brad Thomas, PhD1*, Yanglong Mou, PhD1*, Lauryn Keeler2*, Christophe Magnan, PhD3*, Vincent Funari, PhD4*, Lawrence Weiss, MD4*, Shari Brown, MD5* and Sally Agersborg, M
Slide 2: Disclosure
Disclosure: • All authors of this work are/were employees of NeoGenomics Incorporated. • Funding for this work was entirely by NeoGenomics Incorporated.
Slide 3: Abstract
Abstract • Background: Gene Fusion events are common occurrences in malignancies, and are frequently drivers of malignancy. FISH and qPCR are two methods often used for identifying highly prevalent gene fusions/translocations. However, these are single target assays, requiring a lot of effort and sa
Slide 4: Background
Background • The outstanding need our work is intended to address is for sensitive and specific fusion calling from RNA sequencing (RNAseq) data. • Limiting false-positive and false-negative calls is critical for high-throughput fusion calling. Manual review is time-intensive, and must be assisted b
Slide 5: Methods (1)
Methods (1) • RNAseq data obtained using an in-house developed QIAseq-based panel. • We initially used five distinct fusion-calling pipelines: – Arriba – STAR-Fusion – FusionCatcher – deFuse – kallisto/pizzly • Along with logistic variables – yes/no the fusion caller called the candidate fusion – we
Slide 6: Methods (2)
Methods (2) • The metrics and logistic variables were used as features for modeling • Software for modeling consisted of the R application randomForest which was run within the caret optimization environment. • Training employed the default 500 trees (ntrees) with 10-fold cross- validation. • Data c
Slide 7: Results (1)
Results (1) • In a proof-of-concept study, we compared a random forest classifier with the simpler approach of only requiring that n of the five callers make the call. Random Forest Require 5 callers Require at least 4 Require at least 3 Variable Importance Model callers callers Accuracy (Acc) 0.984
Slide 8: Results (2)
Results (2) • The successful initial proof-of-concept has led to further work. We have improved on modeling with resulting improved sensitivity. Our current model, applied to hematologic malignancies, requires the three independent fusion callers – Arriba – STAR-Fusion – FusionCatcher We have also i
Slide 9: Discussion
Discussion We have arrived at a random forest model that can accurately classify candidate fusions as true or false fusions. Increased sensitivity has brought with it new challenges. Our current focus is on discrimination between driver fusions and “benign” fusions.
Slide 10: References
References • Arriba software https://github.com/suhrig/arriba • STAR-Fusion software https://github.com/STAR-Fusion/STAR-Fusion • FusionCatcher software https://sourceforge.net/projects/fusioncatcher • deFuse software https://github.com/amcpherson/defuse • kallisto software https://pachterlab.github