Ute Neugebauer

Sample size planning for classification models

In biospectroscopy, suitably annotated and statistically independent samples (e.g. patients, batches, etc.) for classifier training and testing are scarce and costly. Learning curves show the model performance as function of the training sample size and can help to determine the sample size needed to train good classifiers. However, building a good model is actually not enough: the performance must also be proven. We discuss learning curves for typical small sample size situations with 5-25 independent samples per class. Although the classification models achieve acceptable performance, the learning curve can be completely masked by the random testing uncertainty due to the equally limited test sample size. In consequence, we determine test sample sizes necessary to achieve reasonable precision in the validation and find that 75-100 samples will usually be needed to test a good but not perfect classifier. Such a data set will then allow refined sample size planning on the basis of the achieved performance. We also demonstrate how to calculate necessary sample sizes in order to show the superiority of one classifier over another: this often requires hundreds of statistically independent test samples or is even theoretically impossible. We demonstrate our findings with a data set of ca. 2550 Raman spectra of single cells (five classes: erythrocytes, leukocytes and three tumour cell lines BT-20, MCF-7 and OCI-AML3) as well as by an extensive simulation that allows precise determination of the actual performance of the models in question.

Identification and differentiation of single cells from peripheral blood by Raman spectroscopic imaging

Medical diagnosis can be improved significantly by fast, highly sensitive and quantitative cell identification from easily accessible body fluids. Prominent examples are disseminated tumor cells circulating in the peripheral blood of cancer patients. These cells are extremely rare and therefore difficult to detect. In this contribution we present the Raman spectroscopic characterization of different cells that can be found in peripheral blood such as leukocytes, leukemic cells and solid tumor cells. Leukocytes were isolated from the peripheral blood from healthy donors. Breast carcinoma derived tumor cells (MCF-7, BT-20) and myeloid leukaemia cells (OCI-AML3) were prepared from cell cultures. Raman images were collected from dried cells on calcium fluoride slides using 785 nm laser excitation. Unsupervised statistical methods (hierarchical cluster analysis and principal component analysis) were used to visualize spectral differences and cluster formation according to the cell type. With the help of supervised statistical methods (support vector machines) a classification model with 99.7% accuracy rates for the differentiation of the cells was built. The model was successfully applied to identify single cells from an independent mixture of cells based on their vibrational spectra. The classification was confirmed by fluorescence staining of the cells after the Raman measurement.

Cell wall investigations utilizing tip-enhanced Raman scattering.

Tip‐enhanced Raman scattering is used to investigate the surface structure of the cell wall of Staphylococcus epidermidis with high lateral resolution and chemical specificity. For most biological samples, the transmission tip‐enhanced Raman scattering set‐up is ideally suited because it allows the specimen to be kept under specific environmental conditions, whereas it most efficiently collects the signal created at the field‐enhancing probe. Special emphasis is given here to the parameters required to reproducibly set up the instrument, such that field‐enhancement factors can be estimated properly. Also the importance of control experiments to avoid misinterpretation of signals will be emphasized by an example.

Toward a Spectroscopic Hemogram: Raman Spectroscopic Differentiation of the Two Most Abundant Leukocytes from Peripheral Blood

The first response to infection in the blood is mediated by leukocytes. As a result crucial information can be gained from a hemogram. Conventional methods such as blood smears and automated sorting procedures are not capable of recording detailed biochemical information of the different leukocytes. In this study, Raman spectroscopy has been applied to investigate the differences between the leukocyte subtypes which have been obtained from healthy donors. Raman imaging was able to visualize the same morphological features as standard staining methods without the need of any label. Unsupervised statistical methods such as principal component analysis and hierarchical cluster analysis were able to separate Raman spectra of the two most abundant leukocytes, the neutrophils and lymphocytes (with a special focus on CD4(+) T-lymphocytes). For the same cells a classification model was built to allow an automated Raman-based differentiation of the cell type in the future. The classification model could achieve an accuracy of 94% in the validation step and could predict the identity of unknown cells from a completely different donor with an accuracy of 81% when using single spectra and with an accuracy of 97% when using the majority vote from all individual spectra of the cell. This marks a promising step toward automated Raman spectroscopic blood analysis which holds the potential not only to assign the numbers of the cells but also to yield important biochemical information.

Light-triggered CO release from nanoporous non-wovens

The water insoluble and photoactive CO releasing molecule dimanganese decacarbonyl (CORM-1) has been non-covalently embedded into poly(l-lactide-co-d/l-lactide) fibers via electrospinning to enable bioavailability and water accessibility of CORM-1. SEM images of the resulting hybrid non-wovens reveal a nanoporous fiber morphology. Slight CO release from the CORM-1 in the electrospinning process induces nanoporosity. IR spectra show the same set of carbonyl bands for the CORM-1 precursor and the non-woven. When the material was exposed to light (365-480 nm), CO release from the incorporated CORM-1 was measured via heterogeneous myoglobin assay, a portable CO electrode and an IR gas cuvette. The CO release rate was wavelength dependent. Irradiation at 365 nm resulted in four times faster release than at 480 nm. 3.4 μmol of CO per mg non-woven can be generated. Mouse fibroblast 3T3 cells were used to show that the hybrid material is non-toxic in the darkness and strongly photocytotoxic when light is applied.

Analysis of Fe(III) Heme Binding to Cysteine-Containing Heme-Regulatory Motifs in Proteins

Regulatory heme binds to specific motifs in proteins and controls a variety of biochemical processes. Several of these proteins were recently shown to form complexes with ferric and/or ferrous heme via a cysteine residue as axial ligand. The objective of this study was to examine the heme-binding properties of a series of cysteine-containing peptides with focus on CP motif sequences. The peptides displayed different binding behavior upon Fe(III) heme application with characteristic wavelength shifts of the Soret band to 370 nm or 420-430 nm and in some cases to both wavelengths. Whereas for most of the peptides containing a cysteine only a shift to 420-430 nm was observed, CP-containing peptides exhibited a preference for a shift to 370 nm. Detailed structural investigation using Raman and NMR spectroscopy on selected representatives revealed different binding modes with respect to iron ion coordination, which reflected the results of the UV-vis studies. A predicted short sequence stretch derived from dipeptidyl peptidase 8 was additionally examined with respect to CP motif binding to heme on the peptide as well as on the protein level. The heme association was confirmed with the first solution structure of a CP-peptide-heme complex and, moreover, an inhibitory effect of Fe(III) heme on the enzymes activity. The relevance of both the use of model compounds to elucidate the molecular mechanism underlying regulatory heme binding and its potential for the investigation of regulatory heme control is discussed.

Charge Isomers of Myelin Basic Protein: Structure and Interactions with Membranes, Nucleotide Analogues, and Calmodulin

As an essential structural protein required for tight compaction of the central nervous system myelin sheath, myelin basic protein (MBP) is one of the candidate autoantigens of the human inflammatory demyelinating disease multiple sclerosis, which is characterized by the active degradation of the myelin sheath. In this work, recombinant murine analogues of the natural C1 and C8 charge components (rmC1 and rmC8), two isoforms of the classic 18.5-kDa MBP, were used as model proteins to get insights into the structure and function of the charge isomers. Various biochemical and biophysical methods such as size exclusion chromatography, calorimetry, surface plasmon resonance, small angle X-ray and neutron scattering, Raman and fluorescence spectroscopy, and conventional as well as synchrotron radiation circular dichroism were used to investigate differences between these two isoforms, both from the structural point of view, and regarding interactions with ligands, including calmodulin (CaM), various detergents, nucleotide analogues, and lipids. Overall, our results provide further proof that rmC8 is deficient both in structure and especially in function, when compared to rmC1. While the CaM binding properties of the two forms are very similar, their interactions with membrane mimics are different. CaM can be used to remove MBP from immobilized lipid monolayers made of synthetic lipids - a phenomenon, which may be of relevance for MBP function and its regulation. Furthermore, using fluorescently labelled nucleotides, we observed binding of ATP and GTP, but not AMP, by MBP; the binding of nucleoside triphosphates was inhibited by the presence of CaM. Together, our results provide important further data on the interactions between MBP and its ligands, and on the differences in the structure and function between MBP charge isomers.

IR Spectroscopic Methods for the Investigation of the CO Release from CORMs

Carbon monoxide (CO) is a toxic gas for mammals, and despite this fact, it is naturally produced in these organisms and has been proven to be beneficial in medical treatments, too. Therefore, CO-releasing molecules (CORMs) are intensively developed to administer and dose CO for physiological applications. Nearly all of these compounds are metal carbonyl complexes, which have been synthesized and investigated. However, for most of these CORMs, the exact reaction mechanisms of CO release is not completely elucidated, although it is of utmost importance. The widely used myoglobin assay for testing the CO release has several disadvantages, and therefore, different methods have to be applied to characterize CORMs. In this work, different setups of IR absorption spectroscopy are used to analyze and quantify the CO release during the decay of various CORMs: IR spectroscopy of the gas phase is applied to follow the CO liberation, and attenuated total reflection (ATR) IR spectroscopy is used to record the decay of the metal carbonyl. IR spectroscopy supported by DFT calculations yields valuable insights in the CO release reaction mechanism. The focus is set on two different CORMs: CORM-2 (Ru2(CO)(6)Cl(4)) and on the photoactive CORM-S1 (photoCORM [Fe(CO)2(SCH2CH2NH2)2]). Our results indicate that the CO liberation from CORM-2 strongly depends on sodium dithionite, which is required for the commonly applied myoglobin assay and that CORM-S1 loses all its bound CO molecules upon irradiation with blue light.

CORM-EDE1: A Highly Water-Soluble and Nontoxic Manganese-Based photoCORM with a Biogenic Ligand Sphere

[Mn(CO)5Br] reacts with cysteamine and 4-amino-thiophenyl with a ratio of 2:3 in refluxing tetrahydrofuran to the complexes of the type [{(OC)3Mn}2(μ-SCH2CH2NH3)3]Br2 (1, CORM-EDE1) and [{(OC)3Mn}2(μ-SC6H4-4-NH3)3]Br2 (2, CORM-EDE2). Compound 2 precipitates during refluxing of the tetrahydrofuran solution as a yellow solid whereas 1 forms a red oil that slowly solidifies. Recrystallization of 2 from water yields the HBr-free complex [{(OC)3Mn}2(μ-S-C6H4-4-NH2)2(μ-SC6H4-4-NH3)] (3). The n-propylthiolate ligand (which is isoelectronic to the bridging thiolate of 1) leads to the formation of the di- and tetranuclear complexes [(OC)4Mn(μ-S-nPr)2]2 and [(OC)3Mn(μ-S-nPr)]4. CORM-EDE1 possesses ideal properties to administer carbon monoxide to biological and medicinal tissues upon irradiation (photoCORM). Isolated crystalline CORM-EDE1 can be handled at ambient and aerobic conditions. This complex is nontoxic, highly soluble in water, and indefinitely stable therein in the absence of air and phosphate buffer. CORM-EDE1 is stable as frozen stock in aqueous solution without any limitations, and these stock solutions maintain their CO release properties. The reducing dithionite does not interact with CORM-EDE1, and therefore, the myoglobin assay represents a valuable tool to study the release kinetics of this photoCORM. After CO liberation, the formation of MnHPO4 in aqueous buffer solution can be verified.

Fast differentiation of SIRS and sepsis from blood plasma of ICU patients using Raman spectroscopy

Currently, there is no biomarker that can reliable distinguish between infectious and non-infectious systemic inflammatory response syndrome (SIRS). However, such a biomarker would be of utmost importance for early identification and stratification of patients at risk to initiate timely and appropriate antibiotic treatment. Within this proof of principle study, the high potential of Raman spectroscopy for the fast differentiation of non-infectious SIRS and sepsis is demonstrated. Blood plasma collected from 70 patients from the intensive care unit (31 patients with sepsis and 39 patients classified with SIRS without infection) was analyzed by means of Raman spectroscopy. A PCA-LDA based classification model was trained with Raman spectra from test samples and yielded for sepsis a sensitivity of 1.0 and specificity of 0.82. These results have been confirmed with an independent dataset (prediction accuracy 80%).