Session TOD. There are 4 abstracts in this session.

Session: Cancer Early Detection and Prevention, time: 3:00 - 3:25 pm

Translational Applications of Mass Spectrometry to Early Cancer Detection and Cancer Treatment

Amanda Paulovich
FHCRC, Seattle, WA


Tips and Tricks (if submitted):

Session: Cancer Early Detection and Prevention, time: 3:25 - 3:50 pm

Helping Bad Biomarkers Fail Fast: Strategies for Prioritizing Candidate Biomarkers

Karin Rodland
Pacific Northwest National Laboratory, Richland, WA

Biomarker discovery strategies often feature heavy reliance on lists of differentially abundant transcripts or proteins derived from unbiased comparisons of diseased versus healthy tissues or biofluids, often with an emphasis on depth of coverage at the expense of statistical power. Translation of candidate biomarkers into clinically approved assays requires relatively high throughput testing on a large number of patients from one or more independent cohorts – and the logistics of doing this with tens to hundreds of markers on many hundreds of patients is daunting, particularly when considering that the typical biomarker failure rate is at least 85%. Here we will describe a strategy based on targeted mass spectrometry, potentially augmented with pathway-specific information, for rapidly and efficiently disqualifying biomarker candidates that are not sufficiently robust for clinical applications.

Tips and Tricks (if submitted):

Session: Cancer Early Detection and Prevention, time: 3:50 - 4:05 pm

Digitizing the Proteomes From Big Tissue Biobanks

Jan Muntel3; Nick Morrice2; Christie Hunter1; Roland Bruderer3; Nicholas Dupuis3; Lukas Reiter3
1SCIEX, Redwood City, CA; 2SCIEX, Warrington, UK; 3Biognosys, Schlieren, Switzerland

Tissue biopsies have been preserved and stored in biobanks for more than a century in the hope that their future analysis will provide a better understanding of health and disease. These samples are often very well characterized by classical pathological methods and provide great potential for precision medicine and the discovery of new diagnostic/stratification markers and therapeutic targets. A powerful way to take advantage of this repository is to quantify large numbers of proteins across all the samples so that correlations can be made with respect to various health and disease states. Such an endeavor would require highly reproducible sample preparation, a robust analytical platform for high throughput sample analysis, as well as robust data analysis.

Here, microflow SWATH Acquisition was used to generate quantitative proteomics data on a cohort of colon cancer samples from a biobank. This study demonstrates how high throughput proteomics can be used to interrogate these precious samples from biobanks and how this research can pave the way to a better understanding of health and disease. 105 colon cancer samples were analyzed in ~5 days with high analytical depth (~4500 proteins quantified across the healthy and disease sample types). Using Spectronaut for data analysis, quantitative results were quickly generated, revealing 3 cancer subtypes and some interesting protein clusters.

Tips and Tricks (if submitted):

Session: Cancer Early Detection and Prevention, time: 4:05 - 4:20 pm

Can AI find a tree in the woods? Issues and considerations for hypothesis free discovery in large clinical proteomics datasets.

David Bramwell; Will Dracup
Biosignatures Ltd, Newcastle, United Kingdom

It is a common paradigm in ‘omics sciences to make many complex measures of a sample and to use AI or statistical analysis to find ‘biomarkers’ that separate specified groups. This is usually termed ‘hypothesis free discovery’ as little or no additional information is used to pre-select candidate analytes.

It is still typical to see ‘biomarker discovery’ experiments that take tens of disease cases and compare them to a matched set of controls. In general, questions like this are ‘ill posed’ i.e. there are far more analytes than case examples to learn from. Without rebalancing the data via analyte pre-selection, many algorithms will be prone to ‘over-learning’ and will not produce an answer that validates on new samples. An additional issue is that diagnosis information is usually binary or ordinal with few cases, which limits the pre-selection tools that can be used.

Clinical prognostics, diagnostics and screens are areas of great need and intense research interest. This presentation enumerates some of these issues on a large clinical proteomics data set. We ask a question that is asked surprisingly infrequently; ‘Can we use our algorithms build a model to measure something that we know is present in the data set?’ To give an example, will the AI systems be able to select the right combination of analytes to measure A1AT from the proteomics data set if we give them the clinical immunoassay A1AT measures to learn from alongside the proteomics data?

Data for several clinical analytes will be presented, some known to be in the proteomics measures and some not expected to be within the measurement limits. The quantifiable issues will be shown and discussed in terms of what these results imply for hypothesis free discovery of disease biomarkers.

Tips and Tricks (if submitted):