22455 research outputs found
Sort by
Assessing lanthanide-dependent methanol dehydrogenase activity and the syntheses of citrate based siderophores
Get PDFKontextabhängiges Prompt- und Flowdesign in Sprachassistenzsystemen
Get PDFWith the rise of Conversational User Interfaces (CUIs), Human-Computer Interaction has made a leap towards natural and intuitive interactions between humans and computers. In the car, CUIs provide a particularly suitable and low-distraction interaction framework. However, inconsiderately designed speech-based interactions with high complexity can result in the opposite effect and increase drivers’ cognitive load. CUI designers for in-car interfaces hence face the challenge of developing voice-based interactions for a potentially vulnerable target group in a demanding setting. At the same time, they are not supported by sufficient as well as sufficiently tailored and empirically validated CUI design guidelines. This thesis closes this gap, by answering the research question of how to context-sensitively design CUI prompts and flows in the car. It does so under consideration of various conversational contexts to adequately account for the multi-facetted nature of human interactions. To this end, five research projects develop and validate concrete linguistic-driven design guidelines for CUI prompts and flows.
To determine an efficient way of validating CUI prompts, a first round of studies was conducted to answer the research question of how to efficiently validate in-car prompts in the paper A Question of Fidelity. Online crowdsourcing studies emerged as a valid alternative to large-scale driving simulator studies. Subsequently, research identified linguistic parameters with a potential impact on the user experience of CUI prompts in How to Design the Perfect Prompt. Three ensuing studies validated the obtained linguistic best practices and prompt design guidelines for different conversational contexts, namely a) the type of interaction, b) the domain of interaction, and c) the initiation of interaction. How to Design the Perfect Prompt, Secure, Comfortable or Functional, and How May I Interrupt showed that CUI prompts need to display a suitable level of (in)formality, complexity/simplicity, and (im)mediacy. Furthermore, an informal, straightforward, and result-oriented speaking style under consideration of the abovementioned contexts is advised. Proactive prompts are thereby specifically dependent on a low level of linguistic complexity as well as a suggestive tone of voice. Additionally, proactive in-car interactions need to carefully consider when to interrupt drivers. To gain insights into best practices for designing CUI flows, the paper Failing With Grace explores error handling strategies from Human-Human-Interaction and their applicability to Human-Computer Interaction. The “Principle of Least Collaborative Effort” and concomitant considerations around so-called costs aid CUI practitioners in designing nuanced user-centric and efficient dialog flows for both successful and erroneous conversations
Stellenwert der ultraschallunterstützten Stanzbiopsie bei der Abklärung einer Lymphadenopathie
Get PDFCharakterisierung unterschiedlicher in vivo und in vitro Modelle akuter myeloischer Leukämie und Evaluation hinsichtlich therapeutischer Anwendung von RIG-I aktivierender, bifunktioneller siRNA
Get PDFErkennung von parodontalem Knochenabbau auf periapikalen Röntgenbildern mittels künstlicher Intelligenz
Get PDFGenauigkeit der Übertragung der räumlichen Position von enossalen Implantaten durch gedruckte und gefräste Bohrschablonen in einem In-vitro Modell
Get PDFValidation of genetically encoded sensors to measure intracellular potassium and metabolism in neurons
Get PDFBiological processes are highly dynamic throughout all levels of organization, ranging from the molecular level up to the full organism. To understand these dynamic processes, they can be observed with an array of different analytical methods, each with particular advantages and limitations. Since their emergence, genetically encoded sensors quickly became the method of choice for many researchers due to their excellent spatial and temporal resolution and the minimally invasive nature of these sensors. In addition, the ever-growing palette of available sensors allows measurements of a wide range of different analytes. However, genetically encoded sensors also have some key limitations, such as sub-optimal affinity, off-target recognition and environmental sensitivity. Therefore, careful interpretation of results obtained with genetically encoded sensors is warranted and potential confounding factors need to be considered in order to avoid artifactual results. Both studies discussed in this thesis address shortcomings of genetically encoded sensors and how to mitigate them.
In study I we characterized lc-LysM-GEPII 1.0, a genetically encoded sensor for potassium ions (K+), with respect to its capacity to resolve neuronal K+ changes upon neuronal activity. lc-LysM-GEPII 1.0 was unable to resolve small K+ dynamics during spontaneous neuronal activity, likely because it might be saturated at physiological K+ concentrations, but reliably detected more pronounced K+ decreases during strong, tetanic activity evoked by application of Bicuculline. We confirmed these results in vivo by fluorescence lifetime imaging of lc-LysM-GEPII 1.0 in the cortex of living mice. We could not observe lifetime changes at baseline, but peri-infarct depolarizations induced by occlusion of the middle cerebral artery led to strong increases in the fluorescence lifetime of lc-LysM-GEPII 1.0. We conclude that lc-LysM-GEPII 1.0 is able to resolve K+ dynamics upon strong neuronal activity but needs to be improved with respect to affinity and dynamic range to measure responses elicited by milder stimulation. To aid this development, we developed an optogenetic stimulation approach that allowed us to titrate the sensitivity of lc-LysM-GEPII 1.0 and will help to compare the performance of different sensor variants.
In study II we developed a novel method to assess the pH sensitivity of genetically encoded sensors without the need for prior purification of the sensor protein. Study II initially aimed to investigate neuronal energy metabolism within the context of GABAergic inhibition. Upon application of GABA to primary cultured neurons, we observed an increase of the FRET ratio of the lactate sensor Laconic, which was mediated by efflux of bicarbonate ions through GABAA receptors, leading to an acidification of the cytosol. While pH changes can lead to artifactual signal changes of genetically encoded sensors, pH can also act as a second messenger eliciting physiological alterations of the levels of the analyte of interest. Therefore, a signal of a genetically encoded sensor can be an artifact, a real change of analyte levels, or a combination of both. We developed a cost-effective and easy-to-use method to separate the pH sensitivity of genetically encoded sensors from the pH-induced physiological analyte response by PFA fixation. This approach, which we call Dead Cell Imaging, preserves sensor fluorescence while stopping all physiological processes. Using this method, we confirmed that the signal change of Laconic upon GABA application is a pH artifact. Furthermore, Dead Cell Imaging provides temporal information about the pH sensitivity of a genetically encoded sensor, which can help to identify complex pH artifacts and is not resolved by canonical methods of addressing the pH sensitivity of a sensor
The theory of Boolean counterfactual reasoning
Get PDFUrsache und Wirkung bilden eine Grundlage dafür, wie Menschen ihre Umwelt wahrnehmen. Wir erklären unsere Umwelt, indem wir Beobachtungen als Wirkung anderer, selbsterklärender Beobachtungen beschreiben. Angenommen, ein Haus brennt nach einem Blitzschlag. In diesem Fall können wir den Blitz als selbsterklärendes Ereignis betrachten, welches das Feuer in dem Haus verursacht. Künstliche Intelligenz hingegen basiert derzeit größtenteils auf dem Konzept der Korrelation. Dies bedeutet, dass sie ausdrücken kann, wie die Beobachtung eines Feuers die Wahrscheinlichkeit für einen Blitzschlag erhöht und umgekehrt. Eine auf Korrelation basierende künstliche Intelligenz kann jedoch den Blitz nicht als Ursache des Feuers erkennen. Während korrelationsbasierte künstliche Intelligenz nur Fragen zur Wahrheit oder Wahrscheinlichkeit von Aussagen beantworten kann, ermöglichen kausale Erklärungen die Beantwortung zweier weiterer Fragetypen: Fragen zu den Folgen externer Interventionen und kontrafaktische Fragen, also Fragen der Form: „Was wäre, wenn...?“. Man kann beispielsweise intervenieren und einen Blitzableiter installieren, um Brände zu verhindern, die durch Blitzschläge verursacht werden. Beobachten wir einen Blitzschlag, der ein Feuer in einem Haus ohne Blitzableiter verursacht, folgt aus dieser kausalen Erklärung, dass kein Feuer ausgebrochen wäre, hätte man einen Blitzableiter installiert.
Menschen beschreiben ihre Umwelt nicht nur in Form kausaler Ursache-Wirkungs-Beziehungen. Oft sind wir auch mit unsicherem Wissen oder mit Beziehungen zwischen Individuen konfrontiert. Wenn wir beispielsweise eine Gruppe von Menschen betrachten, können zwei Individuen innerhalb dieser Gruppe Freunde sein, d.h., sie stehen in Beziehung zueinander. Darüber hinaus kann man annehmen, dass Freunde von Rauchern mit erhöhter Wahrscheinlichkeit ebenfalls rauchen. Wir sind uns also unsicher, ob Freunde von Rauchern zwangsläufig selbst rauchen. Schließlich nehmen wir an, dass die Wahrscheinlichkeit, Raucher zu sein, für alle Personen mit derselben Anzahl von rauchenden Freunden übereinstimmt. Die Annahme, dass Individuen mit denselben Eigenschaften sich gleich verhalten, kann Berechnungen vereinfachen oder den Raum möglicher Beschreibungen eines Datensatzes verkleinern. Das Feld der statistischen relationalen künstlichen Intelligenz kombiniert die prädikatenlogische Repräsentation von Beziehungen innerhalb eines Bereichs mit der Repräsentation von Unsicherheit durch Wahrscheinlichkeiten. Hierbei dient Prädikatenlogik als Sprache, um die Austauschbarkeit von Individuen in einem Modell oder Datensatz auszudrücken. Dieser Ansatz führte unter anderem zu Markov-Logik-Netzwerken und probabilistischer Logikprogrammierung.
Diese Monographie beginnt mit einem Überblick über die zuvor genannten Formalismen und führt in Pearls Kausalitätstheorie ein. Statt jeden Formalismus einzeln um kausales Schließen zu erweitern, betrachtet diese Arbeit einen axiomatischen Ansatz. Sie erweitert Bochmans Theorie der kausalen Logik zu einer Axiomatisierung des deterministischen kausalen Schließens im Boolschen Fall. Diese Axiomatisierung führt dann zur abduktiven Logikprogrammierung als Sprache für kausales Wissen. Anschließend werden abduktive Logikprogramme mit Markov-Logik-Netzwerken zu gewichteten abduktiven Logikprogrammen kombiniert. Auf diese Weise erhält man einen allgemeinen Formalismus für kausales Schließen mit unsicherem Wissen. Die Arbeit spezifiziert die kausalen Fragentypen, d.h. Fragen zu den Folgen externer Interventionen und kontrafaktische Fragen, im Kontext der gewichteten abduktiven Logikprogrammierung. Durch Einbettung verschiedener Formalismen aus dem Feld der statistisch relationalen künstlichen Intelligenz wird das gegebene kausale Schließen übertragen und konsistentes kausales Schließen über verschiedene Formalismen hinweg garantiert. Weiter wird kontrafaktisches Schließen in azyklischen ProbLog-Programmen untersucht. Ein Resultat gibt Regularitätsbedingungen, unter denen azyklische ProbLog-Programme aus ihrer kontrafaktischen Ausgabe rekonstruiert werden können. ProbLog scheint daher als Sprache für kontrafaktisches Schließen besonders geeignet zu sein. Die Lerntheorie von ProbLog-Programmen basiert jedoch auf statistischen Tests, die keine Informationen über den der Daten zugrunde liegenden kausalen Mechanismus liefern. Somit ist es unzulässig, die resultierenden Programme für kontrafaktisches Schließen zu verwenden. Im letzten Teil wird ein Fragment beschrieben, in welchem sich ProbLog-Programme aus ihrer Verteilung rekonstruieren lassen. Weiter wird argumentiert, dass dieses Fragment einen Ansatz liefert, um ProbLog-Programme für kontrafaktisches Schließen aus Daten zu lernen.Human reasoning is heavily based on distinguishing causes from their effects. We typically understand our environment by explaining observations, i.e., effects, with the help of self-evident a priori knowledge, i.e., causes. Assume, for instance, that a lightning strike hits a house and the house burns down. In this case, we consider the lightning as self-evident, a priori knowledge, which explains the fire in the house. This leads us to the judgment: "The lightning caused the fire in the house''.
Unfortunately, nowadays artificial intelligence is mostly built on the concept of correlation. In our example, this means we can express that observing a fire increases the probability of a lightning strike hitting our house, and vice versa. However, a correlation-based artificial intelligence cannot recognize the lightning as the fire's cause. While correlation-based artificial intelligence only supports queries about the truth or probabilities of statements, causal explanations allow two additional query types: queries for the effects of external interventions and counterfactual queries, i.e., queries of the form: "What if...?''. In our scenario, we can, for instance, intervene and install a conductor to prevent fires being caused, i.e. explained by lightning strikes. Furthermore, if we observe a lightning strike hitting our house followed by a fire breaking out, we may conclude from a causal explanation that the fire would not have occurred had we installed a lightning rod beforehand.
Humans reason not only on causes, but also on uncertainties, represented by probabilities, and on relations between components or individuals of a given domain formalized in logic, preferably first-order logic. When, for example, given a group of people, two individuals within this group may be friends, i.e., they share a relationship. Moreover, one may assume that friends of smokers are more likely to be themselves smokers, meaning that we are uncertain about the implication whether friends of smokers necessarily smoke. Finally, we may assume that all individuals with the same number of smoking friends are equally likely to smoke themselves. The assumption that individuals with the same properties behave in the same way may speed up calculations or shrink the space of possible descriptions for a dataset. The field of statistical relational artificial intelligence combines first-order logic reasoning on relations in a domain and probabilistic reasoning on uncertainty, whereby first-order logic serves as a language for expressing the interchangeability of individuals in a probabilistic model or dataset. Among other things, this approach has so far resulted in Markov logic networks and probabilistic logic programs.
The work reported in this monograph first provides a brief overview of the aforementioned formalisms and introduces Pearl's causal reasoning. Instead of continuously expanding every distinct formalism with causal reasoning, the thesis proposes an axiomatic approach to Boolean causal reasoning under uncertainty. Initially, we extend Bochman's logical theory of causality to obtain a complete axiomatization of deterministic Boolean causal reasoning, leading to abductive logic programming. We then generalize abductive logic programming to the framework of Markov logic networks, which employs a maximum entropy approach extending first-order logic to handle uncertainty. In this way, we obtain weighted abductive logic programming as a general framework tailored to Boolean causal reasoning under uncertainty. We specify the causal query types, i.e., queries for the effects of external interventions and counterfactual queries, within the context of weighted abductive logic programming. By embedding a formalism into our framework and subsequently transferring the causal reasoning there, we derive causal reasoning in widespread formalisms of statistical relational artificial intelligence.
In particular, our approach guarantees consistent causal reasoning across these frameworks. Next, we focus on counterfactual reasoning in acyclic ProbLog programs, where the proposed approach is implemented in Kiesel's WhatIf solver. We show that sufficiently well-behaved ProbLog programs can be reconstructed from their counterfactual output. Hence, ProbLog is a particularly suitable framework to address counterfactual reasoning. However, learning ProbLog programs from data is usually based on statistical tests, lacking information about the underlying causal mechanism. This makes it infeasible to use the resulting programs for counterfactual reasoning. To address this issue in the last part, we propose a fragment of ProbLog that allows reconstructing a program from its induced distribution, finally enabling us to learn programs supporting counterfactual queries
Statistical approaches for modeling network and public health data
Get PDFThe rapid technological advancement that characterized the past few decades has brought about an increasingly large amount and variety of data. This wealth of data naturally comes with further complexity, thus requiring increasingly sophisticated and efficient methodologies to extract valuable information from it. In this context, statistical models can serve as effective tools to obtain interpretable insight from the data while adequately quantifying and accounting for the underlying uncertainty. This thesis deals with the statistical modeling of two broad data categories that are prominent in modern times: network data and public health data. After an introductory Part I, the thesis comprises a total of eleven contributions, which can be divided into three further parts.
Part II, composed of four contributions, deals with the statistical analysis of network data. Networks can broadly be defined as groups of interconnected people or things. This thesis focuses mostly on social and economic networks, and on statistical models aimed at capturing and ex- plaining the mechanisms leading to the formation of ties between actors within the network. We specifically concern ourselves with two broad model families, namely latent variable models and exponential random graph models. The first two contributions in this section introduce and compare several models from these classes, and showcase them by applying them to real-world network data. The following two contributions extend and apply these models to answer substantive questions in the social sciences. More specifically, the third contribution extends exponential random graph models to deal with the modeling of a massive dynamic bipartite network of patents and inventors to explore the drivers of innovation, while the fourth one uses latent distance models to map the network of popular Twitter users discussing the COVID-19 pandemic, with the goal of investigating polarization on the platform.
Part III, which also comprises four contributions, addresses statistical challenges related to the real-time monitoring and modeling of public health data. More specifically, the chapter tackles questions that emerged during the early stages of the COVID-19 pandemic, mainly by adapting and extending the class of generalized additive mixed models (GAMMs). The fifth contribution develops a statistical model using reported fatal infections data to predict how many of the registered infections will turn out to be lethal in the near future, thereby enabling to effectively monitor the current state of the pandemic. The sixth contribution instead focuses on all reported infections, and proposes a model to nowcast locally detected (but not yet centrally reported) cases by accounting for expected reporting delays, as well as to forecast infections at the regional level in the near future. The seventh contribution proposes a statistical tool to study the dynamics of the case-detection ratio over time, allowing for comparisons of infection figures between different pandemic phases. The chapter is concluded by the eighth contribution, which further demonstrates the effectiveness of GAMMs by applying them to three relevant pandemic-related issues, i.e. the interdependence among infections in different age groups among school children, the nowcasting of COVID-19 related hospitalizations, and the modeling of the weekly occupancy of intensive care units.
Finally, Part IV, composed of three contributions, focuses on the principled estimation of excess mortality, which can generally be defined as the number of deaths from all causes during a crisis beyond what would have been expected had the crisis not occurred. More specifically, the ninth contribution develops a point-estimation method by deploying a corrected version of classical life tables to calculate age-adjusted excess mortality, and applies it to obtain estimates the first year of the COVID-19 pandemic (i.e. 2020) in Germany. The tenth contribution applies the same method to provide updated age-specific estimates for 2021. Finally, the eleventh contribution extends the method to incorporate uncertainty quantification, and deploys it at a broader scale to obtain estimates for 30 developed countries in the first two years of the COVID-19 crisis. The results are further compared with existing estimates published in other major scientific outlets, highlighting the importance of proper age adjustment to obtain unbiased figures
Development of barcoded GPCR assays to assess the effects of neuroleptics for brain disorders in living cells
Get PDFBackground: G protein-coupled receptors (GPCRs) represent the largest protein family in the human genome and are targeted by 35% of FDA-approved drugs. Their signalling pathways are complex and attractive for drug discovery. GPCR activation can be measured using the split TEV protein-protein interaction technique at the membrane. In addition, pathway reporter gene assays can be used at the transcriptional level via cis-regulatory elements, as these can be activated independently of the transducer pathways of GPCR activation.
Methods: First, I identified whether the combinations of the signal peptide (SP) and the β-arrestin binding motif (i.e., C-terminal tail of the vasopressin 2 receptor (AVPR2); V2R tail) for six GPCRs and cell backgrounds of four cell types affect the β-arrestin recruitments with the split TEV GPCR assay in a transfection-based way. Then I developed a living cell-based plat-form for the multiplex profiling GPCR signalling activities at the transcriptional level, called the barcoded GPCR pathway assay. GPCRs were stably integrated into HEK293 cells to avoid the variability associated with transient transfection. To monitor the multiplex profiling of G protein-coupled receptor (GPCR) signalling activities, a lentivirus-based sensor library was used encoding various genetically encoded pathway sensors, based on cis-regulatory ele-ments.
Results: The split TEV GPCR assays revealed that the combinations of SP and V2R tail have different effects depending on the GPCR, whereas HEK293 cells provided the best perfor-mance in most situations. Then I selected HEK293 cells for the barcoded GPCR pathway as-say. The robustness of the barcoded GPCR pathway assay was confirmed by testing various plate formats, MOIs (multiplicity of infection) of the lentiviral sensor library, and ligand stimula-tion. Furthermore, the reproducibility of the barcoded GPCR pathway assay was demonstrated through seven independent repetitions performed on different days. However, the combination of stable cells and the lentiviral sensor library produced two issues. Firstly, there was crosstalk between different stable cells due to barcode sequencing leakage. Secondly, the next-generation sequencing procedure can result in so-called read eating effects. This occurred because the barcodes from the index sensors, which were introduced when the stable cells were established, have much higher mRNA levels than the barcodes from the lentiviral sensor library. As a result, sequencing resources are intensively used for these frequently occurring barcodes, leaving only a limited and insufficient amount of sequencing resources for low oc-curring barcodes.
Outlook: Although the above two issues need to be addressed, the success of the project is underlined by the robustness and reproducibility of the barcoded GPCR pathway assay, providing a solid foundation for its further development.Hintergrund: G-Protein gekoppelte Rezeptoren (GPCRs) sind die größte Proteinfamilie im menschlichen Genom und das Target von 35% der von der FDA zugelassenen Medikamente. Die Signalwege, die von GPCRs aktiviert werden, sind komplex und zugleich attraktiv für die pharmazeutische Forschung. Die Aktivierung von GPCRs kann einerseits mit der Split-TEV-Protein-Protein-Interaktionstechnik direkt an der Zellmembran gemessen werden. Darüber hinaus können Signalweg-Assays, die auf einem transkriptionellem Readout beruhen und cis-regulierende Elemente nutzen, verwendet werden, da diese Reportergen-Assays unabhängig von den aktivierten Signalwegen GPCR-Aktivitäten abbilden können.
Methoden: Zunächst untersuchte ich, ob die Kombinationen des Signalpeptids (SP) und des β-Arrestin-Bindungsmotivs (d.h. C-terminaler Tail des Vasopressin-2-Rezeptors (AVPR2); V2R-Tail) für sechs GPCRs und in vier verschiedenen Zelltypen die β-Arrestin-Rekrutierung mit dem Split-TEV-GPCR-Assay auf transfektionsbasierte Weise beeinflussen. Anschließend entwickelte ich eine auf lebenden Zellen basierende Plattform für die multiparametrische Profilierung von GPCR-Signalaktivitäten auf transkriptioneller Ebene, den sogenannten Barcoded GPCR Pathway Assay. Die GPCRs wurden einzeln in HEK293-Zellen integriert, um die mit der transienten Transfektion verbundene Variabilität zu vermeiden. Um die Profilierung der Signalaktivitäten von G-Protein gekoppelten Rezeptoren (GPCRs) zu messen, wurde eine Lentivirus-basierte Sensorbibliothek mit verschiedenen genetischen Signalwegsensoren infiziert.
Ergebnis: Die Split-TEV-GPCR-Assays zeigten, dass die Kombinationen von SP und V2R-Schwanz je nach GPCR unterschiedliche Effekte hatten, wobei HEK293-Zellen in den meisten Situationen die beste Assay-Performance zeigten. Daher wählte ich HEK293-Zellen für den Test mit dem multiparametrischen GPCR-Signalweg-Assay aus. Die Robustheit des dieses Assays wurde durch das Testen verschiedener Plattenformate, MOIs (Multiplicity of Infection) der lentiviralen Sensorbibliothek und der Stimulation mit verschiedenen Liganden bestätigt. Darüber hinaus wurde die Reproduzierbarkeit des multiparametrischen GPCR-Signalweg-Assays durch sieben unabhängige Wiederholungen an verschiedenen Tagen nachgewiesen. Die Kombination von stabilen Zellen und der lentiviralen Sensorbibliothek war jedoch mit zwei Problemen verbunden. Zum einen kam es aufgrund von Crosstalk in der Barcode-Sequenzierung zu Überlappungen zwischen verschiedenen stabilen Zellen. Zweitens kann die NGS-Sequenzierungstechnik zu sogenannten Read-Eating-Effekten führen. Diese treten auf, wenn die Barcodes der Index-Sensoren, die bei der Etablierung der stabilen Zellen eingeführt wurden, wesentlich höhere mRNA-Werte aufweisen als die Barcodes der lentiviralen Sensorbibliothek. Dies hat zur Folge, dass die Sequenzierressourcen für diese häufig vorkommenden Barcodes intensiv genutzt werden und nur eine begrenzte und unzureichende Menge an Sequenzierressourcen für die selten vorkommenden Barcodes übrig bleibt.
Ausblick: Obwohl die beiden oben genannten Probleme noch gelöst werden müssen, wird der Erfolg des Projekts durch die Robustheit der Sequenzierbibliothek gewährleistet