Accelerating the Implementation of Technology in the Clinical Lab



Palm Springs, Mar 29 - Apr 3


Educational Grants supported in part by:
Brian Kelly, Danyel Tacker, Andy Hoofnagle, Brian Rappold, Russell Grant, David Herold

Plenary & Keynote Lecture Series


Plenary Lectures

>> Tuesday 14:15 in Rm 4-6

Michael Angelo
Stanford Bio-X

>> Thursday 16:45 in Rm 4-6
Brain Organoids as a Model System for Human Neurodevelopment and Evolution
Alysson Muotri
University of California, San Diego

Structural and transcriptional changes during early brain maturation follow fixed developmental programs defined by genetics. However, whether this is true for functional network activity remains unknown, primarily due to experimental inaccessibility of the initial stages of the living human brain. We developed cortical organoids that spontaneously display periodic and regular oscillatory network events that are dependent on glutamatergic and GABAergic signaling. These nested oscillations exhibit cross-frequency coupling, proposed to coordinate neuronal computation and communication. As evidence of potential network maturation, oscillatory activity subsequently transitioned to more spatiotemporally irregular patterns, capturing features observed in preterm human electroencephalography (EEG). These results show that the development of structured network activity in the human neocortex may follow stable genetic programming, even in the absence of external or subcortical inputs. Our approach provides novel opportunities for investigating and manipulating the role of network activity in the developing human cortex. Applications for neurodevelopmental disorders and brain evolution will be discussed.

Keynote Lectures

Michael Lassman

Tom Rosano
Albany Medical College

Chip Cody
Mcr and VCA – Two R Packages to Facilitate Your Method Comparison and Precision Studies
Andrea Geistanger

Trueness and precision are the key quality attributes of a diagnostic assay and have to be proven in validation experiments throughout each assay development. CLSI does also acknowledge the importance of these criteria, having two guidelines in place, EP9 for method comparison, and EP5 for precision studies describing the design and the analysis of the corresponding experiments. The statistical methodology for both experiments is quite advanced and cannot be operated in a bread and butter software such as Excel. For method comparison studies a Deming regression is required and in some cases also a robust Passing-Bablok regression is state-of-the-art. Classical linear regression methods are not appropriate here, as measurement errors occur for both measurement methods. For precision studies, an appropriate variance-components design should be used and statistically analyzed accordingly.

The mcr R-package is a free available open source R package, which incorporates all analysis methods for method comparison studies, with special focus on the regression methods as Deming or Passing-Bablok regression.

The VCA package is the pendant for precision experiments, where different measurement designs can be analyzed. It is also freely available as open source R-package. Both R packages have been developed and are maintained by the Roche Diagnostics R&D biostatistics department.
The talk will cover the major aspects of the analysis requirements for method comparison and variance-components studies. In addition, we show the features of both R packages, their calculation capabilities as well as the graphical representation possibilities.
Spatial Metabolomics: From Big Data to Single Cells
Theodore Alexandrov
European Molecular Biology Laboratory - Heidelberg

Recent discoveries put metabolism into the spotlight. Metabolism not only fuels cells but also plays key roles in health and disease in particular in cancer, inflammation, and immunity. In parallel, emerging single-cell technologies opened a new world of heterogeneous cell types and states previously hidden beneath population averages. Yet, methods for discovering links between metabolism, cell states, metabolic plasticity and reprogramming on the single-cell level and in situ are crucially lacking. Our research aims to bridge this gap. First, I will explain how the emerging technology of imaging mass spectrometry can be used for the spatial profiling of metabolites, lipids, and drugs in tissues. I will present our cloud and Artificial Intelligence-powered platform METASPACE which is increasingly used across the world. In the second part of my talk I will focus on our method SpaceM for spatial single-cell metabolomics in situ. We applied SpaceM to investigate hepatocytes stimulated with fatty acids and cytokines, a model mimicking the inflammation-associated transition from the fatty liver disease NAFLD to steatohepatitis NASH. We characterized the metabolic state of steatotic hepatocytes and metabolic plasticity associated with the inflammation. We discovered that steatosis and proliferation take place in distinct cell subpopulations, each with a characteristic spatial organization and metabolic signatures. Overall, such methods open novel avenues for understanding metabolism in tissues and cell cultures on the single-cell level.
Mapping the Chemical Space of Biological Systems via MALDI Mass Spectrometric Imaging and <i>in situ</i> Molecular Analysis
Lingjun Li
University of Wisconsin-Madison
Mass spectrometric imaging (MSI) provides an attractive opportunity to detect and probe the molecular content of tissues in an anatomical context. This technique creates distribution maps of select compounds without the need for priori knowledge of target analytes. In this presentation, I will describe our efforts and recent progress in mapping and imaging of a wide variety of signaling molecules in several biological systems, highlighting the unique challenges and important roles of MSI in the areas of proteomics, peptidomics, and metabolomics.

Although high resolution accurate mass (HRAM) MSI platform offers unique advantages for mapping small molecule metabolites due to its high resolution and accuracy measurement, typical MALDI-LTQ-Orbitrap platform suffers from limited utility for large peptide and protein analysis due to its maximum m/z 4000. To overcome this challenge, we employed volatile matrices to produce multiply charged ions in MALDI source via laserspray ionization (LSI) and matrix assisted ionization in vacuum (MAIV) techniques on the MALDI Orbitrap platform. These new ionization techniques enabled substantial expansion of the mass range of the instrument and generated improved fragmentation efficiency compared to traditional MALDI-MS. To further enhance the chemical information extracted from in situ MALDI MSI experiments, we report on a multiplex-MSI method, which combines HRAM MSI technology with data dependent acquisition (DDA) tandem MS analysis in a single experiment. To improve the dynamic range and efficiency of in situ DDA, we introduce a novel gas-phase fractionation strategy prior to MS/MS scans, to decrease molecular complexity of tissue samples for enhanced peptidome coverage. In addition, the application of HRAM MALDI MSI to lipid analysis in a restenosis rat model and the utility of a novel subatmospheric pressure (SubAP)/MALDI source coupled with a Q Exactive HF hybrid quadrupole-orbitrap mass spectrometer for in situ imaging of glycans from formalin-fixed paraffin-embedded (FFPE) tissue sections and its translation to clinical cancer tissue microarray analysis will be highlighted. Finally, to further improve the sensitivity of MALDI MSI, a photoactive compound, 2-nitrobenzaldehyde is used to initiate a nanosecond photochemical reaction (nsPCR). This nsPCR strategy enables enhanced neuropeptide identification and visualization from complex tissue samples through on-demand removal of surrounding matrices within nanoseconds. The utility of this new approach for in situ analysis of endogenous biomolecules is evaluated and demonstrated.
Multiple Instance Learning for Classification in Mass Spectrometry Imaging
Olga Vitek
Northeastern University
Introduction: Mass spectrometry imaging (MSI) has the potential to become a rapid routine analysis tool for detecting tissue conditions such as tumors. Currently, the tumor or non-tumor labels for classification algorithm training are usually obtained by a pathological examination of sub-areas of the tissues. However, due to high within-tumor heterogeneity, and to much higher spatial resolution of MSI than pathological labeling, a pixel profiled by MSI from a tumor-labeled tissue may not contain tumor cells. This undermines the accuracy of the machine learning algorithms.
Objectives: The objective is to improve the classification accuracy with presence of mis-labeled normal pixels in tumor tissues.

Methods: we developed a classification approach based on multiple instance learning (MIL) and convolutional neural network (CNN), that we call mi-CNN. We assume that tumor tissue contains at least one tumor pixel and may contain normal pixels while normal tissue contains no tumor pixel. Unlike in a typical image analysis where the convolution is applicable in the image space, here we apply it to the m/z space to reduce the dimensionality of the potential predictors and to capture the dependencies of m/z features. An expectation-maximization-like algorithm is applied to train pixel-level classification: CNN is trained given the current labels of pixels and pixel labels are imputed by the results of CNN output in each iteration. For image-level classification, the outputs of the last CNN layer are aggregated by a MIL pooling layer to achieve image-level prediction.

Results: To evaluate our model on a dataset with known ground truth, we simulated a dataset of 8 tumor samples and 8 normal samples and ~300 pixels per sample. In tumor samples, only half of the pixels are simulated as tumor pixels. In order to make the simulation as realistic as possible, we added artificial peaks that distinguish the tumor and normal pixels to mass spectra from an MSI investigation of normal tissues. We applied both non-MIL classification methods, specifically Support Vector Machine (SVM) and CNN, as well as MIL classification methods mi-SVM and mi-CNN. The results show that mi-CNN achieves the highest prediction accuracy of 98.3% among these methods. mi-SVM and mi-CNN outperforms their corresponding non-MIL method SVM and CNN.
We also compared mi-CNN and mi-SVM on a human renal cell carcinoma dataset with 16 tissues profiled by desorption electrospray ionization (DESI)-MSI. Two cancer and two normal tissues were randomly selected out of eight cancer tissues and eight normal tissues as testing data. The results show that both mi-CNN and mi-SVM predict some parts in tumor tissues as normal, while mi-CNN predicts much few pixels in normal tissues as tumor than baseline mi-SVM does.

Conclusions: Mi-CNN can capture the dependencies among m/z and distinguish normal pixels in a tumor-labeled tissue, thus improve the classification accuracy.