Raf Van de Plas (Presenter)
Delft University of Technology
Bio: Raf Van de Plas is assistant professor at the Delft University of Technology (TU Delft) in the Netherlands. At the Delft Center for Systems and Control, he works on computational analysis of molecular imaging modalities such as imaging mass spectrometry and microscopy. He holds an adjunct assistant professor position in biochemistry at Vanderbilt University School of Medicine in Nashville, TN. He holds a PhD in Engineering (KU Leuven, 2010), an MSc in Industrial Engineering (Group T, 2002), and an MSc in Artificial Intelligence (KU Leuven, 2003). Before his appointment at TU Delft, he held a research faculty position at the Mass Spectrometry Research Center at Vanderbilt University with Dr. Richard Caprioli. His research focuses on the interface between (i) mathematical engineering and machine learning; (ii) analytical chemistry and instrumentation; and (iii) life sciences and medicine.
Authorship: Raf Van de Plas (1)(2), Jeffrey Spraggins (2), Boone M. Prentice (2), Junhai Yang (2), Richard M. Caprioli (2)
(1) Delft Center for Systems and Control, Delft University of Technology, Delft, The Netherlands. (2) Mass Spectrometry Research Center and the Department of Biochemistry, Vanderbilt University School of Medicine, Nashville, TN, USA.
Clinical pathology employs a broad array of targeted microscopy stains to drive diagnosis. While commonly employed, their labeled nature makes stains inherently targeted, yielding only narrow insight into underlying patho-mechanisms. Imaging MS (IMS) does not require labeling and can map thousands of chemical species in a single experiment, providing an excellent imaging modality for exploratory tissue analysis and discovery of disease markers. Here, we employ machine learning techniques to automatically learn correlative relationships between microscopy and biomolecular species reported by IMS. The ability to empirically capture microscopy-IMS relationships in mathematical models, and open up those models for biological interpretation, enables a novel link between medical practice and exploratory biology and has potential as a new multi-modal pipeline for discovery of clinical markers.
Clinical pathology employs a broad array of targeted microscopy stains to drive prognosis and diagnosis. While these approaches are typically well-understood by medical professionals, the labeled nature of stains makes this type of imaging inherently targeted, yielding only a narrow insight into the underlying patho-mechanism. Imaging mass spectrometry (IMS) does not require chemical labeling and can map thousands of chemical species throughout a tissue section in a single experiment, making it an excellent imaging modality for exploratory tissue analysis and discovery of novel disease markers. While the massively multivariate nature of IMS brings a broader view into the molecular interplay underlying disease, its findings are currently much less connected to the clinical practice, hampering its adoption for diagnosis tasks. In this work, we employ machine learning techniques to automatically learn potential correlative, anti-correlative, and multivariate linear relationships between microscopy-reported target species and protein, peptide, lipid, metabolite, and drug species reported by IMS. The ability to empirically discover and capture microscopy-IMS relationships in mathematical (white-box) models, and the opportunity to open up those models for subsequent biological interpretation, enables a novel link between the medical practice and exploratory biology and has potential as a new multi-modal pipeline for discovery of clinical markers. The approach can be particularly helpful in areas where the established microscopy stains provide limited sensitivity, specificity, or insufficient diagnostic power to discern different disease sub-classes.
The discovery methodology and its implementation extend a multi-modal framework built originally for purposes of data-driven multi-modal fusion (Van de Plas et al., Nature Methods, 2015.). However, fusion-focused applications such as spatial sharpening, out-of-sample prediction, and multi-modal de-noising seek to optimize their cross-modal models in terms of predictive capability, which tends to lead to predictively very capable but complex multivariate models that are difficult to interpret (gray and nearly black-box models). In the case of automated discovery of inter-modal relationships between microscopy stains and IMS, specifically for purposes of subsequent human interpretation of these relationships, we keep the multivariate cross-modal models relatively sparse (white-box models), allowing only the strongest connections between microscopy and IMS to be captured. This entails sacrificing on cross-modal predictive capability in order to obtain an easily interpretable linear combination of ions that correlate or anti-correlate to established microscopy markers.
We demonstrate the utility of this methodology in several case studies, including the characterization of human pancreas tissue in the context of diabetes. In this case study, we are specifically interested in automatically identifying, amongst hundreds of ions, which ion species have a strong relationship to beta cells exclusively. Furthermore, we need to accomplish this microscopy-IMS discovery step with only limited IMS spatial resolution, which by itself is insufficient to differentiate islet substructure. The automated discovery process combines information from MALDI-TOF IMS measurements, acquired in the protein range at a spatial resolution of 15 μm, with immunohistochemically (IHC) stained microscopy measurements at 3.75 μm resolution. The IMS measurements were obtained on a Bruker AutofleX MALDI TOF mass spectrometer in linear mode (mass range: 1000 Da to 20 kDa; lateral resolution: 15 μm). The stained microscopy consists of different IHC stains applied to the pancreas tissue sample: (a) a glucagon stain to label alpha islet cells; (b) an insulin stain to label beta islet cells; (c) a DAPI stain to label cell nuclei, employed as a counter stain; and (d) a somatostatin stain. As an example, the empirically learned microscopy-IMS model is used to automatically identify which ions (out of hundreds) demonstrate (exclusively) a correlative, anti-correlative, or multivariate linear relationship to the beta cells within the Islets of Langerhans, effectively matching and exceeding what can be gleaned from the microscopy stains in isolation.
Conclusions & Discussion
The multivariate regression engine driving the discovery process does not require prior knowledge and is capable of revealing complex multivariate localizations and inter-modality relationships, effectively tying exploratory observations made in IMS to patterns observed via established assays in pathology and the clinic. The automated discovery process for microscopy-IMS relationships can be applied to any microscopy stains accepted in the medical and pathology practice, including IHC stains and FFPE-based samples, and can target specific cell types or anatomical regions (even when the IMS spatial resolution is insufficient to clearly differentiate). It can, for example, answer (in an automated way) questions such as ”Which ions strongly co-localize with Stain A, except in areas where stain B is present as well, and are also tissue-wide absent from Stain C locations”. It can do this without any prior knowledge or specification. These capabilities can be used to help answer specific biological questions and drive subsequent hypothesis building. This automated discovery process can furthermore be a crucial component to exhaustively identifying the connections between clinical microscopy and exploratory IMS.
References & Acknowledgements:
Van de Plas R., Yang J., Spraggins J., and Caprioli R.M., Image fusion of mass spectrometry and microscopy: a multimodality paradigm for molecular tissue mapping. Nature methods, 2015, 12(4), pp.366-372.
IP Royalty: no
|Planning to mention or discuss specific products or technology of the company(ies) listed above:||