119092 research outputs found
Sort by
Cell adhesion and local cytokine control on protein‐functionalized PNIPAM‐co‐AAc hydrogel microcarriers
No full textAchieving adequate cell densities remains a major challenge in establishing economic biotechnological and biomedical processes. A possible remedy is microcarrier‐based cultivation in stirred‐tank bioreactors (STBR), which offers a high surface‐to‐volume ratio, appropriate process control, and scalability. However, despite their potential, commercial microcarriers are currently limited to material systems featuring unnatural mechanical properties and low adaptability. Because matrix stiffness and ligand presentation impact phenotypical attributes, differentiation potential, and genetic stability, biotechnological processes can significantly benefit from microcarrier systems tailorable toward cell‐type specific requirements. This study introduces hydrogel particles co‐polymerized from poly(N‐isopropylacrylamide) (PNIPAM) and acrylic acid (AAc) as a platform technology for cell expansion. The resulting microcarriers exhibit an adjustable extracellular matrix‐like softness, an adaptable gel charge, and functional carboxyl groups, allowing electrostatic and covalent coupling of cell adhesive and cell fate‐modulating proteins. These features enable the attachment and growth of L929 mouse fibroblast cells in static microtiter plates and dynamic STBR cultivations while also providing vital growth factors, such as interleukin‐3, to myeloblast‐like 32D cells over 20 days of cultivation. The study explores the effects of different educt compositions on cell‐particle interactions and reveals that PNIPAM‐co‐AAc microcarriers can provide both covalently coupled and diffusively released cytokine to adjacent cells
From a P‐bridging phosphaketene to μ‐phosphinidenide and μ‐diphosphaurea units at a dinickel core
No full textSalt metathesis of dinickel(II) complex LNi₂Br (1; L is a dinucleating pyrazolate ligand with two β‐diketiminato chelate arms) with Na(OCP) ⋅ (dioxane)₂.₅ yielded LNi₂(PCO) (2) with a P‐bridging phosphaethynolate. Further reaction of 2 with benzyl isocyanide or with an N‐heterocyclic carbene (NHC) triggered decarbonylation and gave LNi₂(PCN‐CH₂Ph) (3) and LNi₂P(NHC) (4) with P‐bridging cyanophosphide and NHC‐phosphinidenide, respectively. Electronic structure analysis indicated a μ₂‐η² : η¹ binding mode of the PCO⁻ anion between the two NiII ions in 2, which is even more pronounced for the [PCN(−CH₂Ph)]⁻ anion in 3. DFT assessment of the formation mechanism of 4 showed that attack at the phosphaketene‐C atom is kinetically preferred but reversible and unproductive, while kinetically more demanding back‐side SN2 attack at the phosphaketene‐P atom triggers CO release with 4 as thermodynamic product. Nucleophilic addition at the phosphaketene‐C could be demonstrated by the strongly exergonic reaction of 2 with KPPh₂, giving unstable K[LNi₂(P(O)CPPh₂)] (5) with a P‐bridging and K⁺‐stabilized diphosphaurea derivative. All new complexes 2–5 have been comprehensively characterized, including by X‐ray diffraction
Two Optimization Approaches for a Small‐Scale Power‐to‐Ammonia Cycle
No full textAmmonia is a promising carbon‐free energy vector. Small‐scale renewable power‐to‐ammonia (P2A) is particularly suited for isolated agricultural areas where ammonia can be used as fuel and fertilizer. This work compares two approaches to simulate and optimize the steady‐state behavior of a novel small‐scale P2A process: Aspen Plus® and MOSAIC®. Aspen Plus® is a commercial flow sheeting software whereas MOSAIC® is a freeware where equations and thermochemical properties need to be specified by hand. It can be shown that the results of MOSAIC® and Aspen Plus® are qualitatively comparable, but not identical. This suggests that the model in MOSAIC® can be improved further, starting with the implementation of a more accurate numeric reactor kinetics and equation of state
Controlling temporally and spatially homogeneous temperature distribution of paper substrates by biogenic phase change hybrid material coatings
No full textHere the performance of phase change material (PCM)‐coated paper made from unbleached kraft pulp is introduced. The applied PCM consists of a mixture of ethylene glycol distearate (EGDS), a well‐known PCM wax material, and a fully substituted cellulose stearoyl ester (CSE). Transfer of the PCM material onto/into paper is achieved by spray as well as blade coating of EGDS + CSE mixture. It is shown that the kind of coating method used does not interfere with observed PCM properties. The significantly higher melt viscosity of the EGDS + CSE blends ensures that the EGDS wax is not bleeding out of the paper, which avoids the use of further encapsulation processes. The PCM behavior, as observed by thermal load measurements, and the thermal buffering of the coated paper is a function of the applied mass of the PCM material applied. The thermal retention exhibited a quasi‐isothermal behavior at ≈65 °C with EGDS + CSE coatings. These effects can offset fluctuations in temperature, and the PCM papers can be employed to achieve a more uniform temperature setting. PCM‐modified papers are therefore interesting candidates for paper‐based packaging or for use in paper‐based sensors, where overheating can strongly affect reliability of results
Enhancing Robot Manipulation Skills through Learning
No full textRobotic manipulation is a core area in robotics that focuses on developing systems that enable robots to interact and modify their environment by manipulating objects in the physical world. From stacking groceries on supermarket shelves to solving difficult assembly tasks in industrial settings, robots are expected to handle multiple manipulation challenges in different environments and handle various objects. Manipulation systems are composed of several modules, such as perception, motion planning, and control, each specialized in solving a part of the task, enabling researchers to focus on specific problems and advance the field in a modular way. However, these modules are still nowadays heavily-engineered to solve a specific problem in a static environment or handle particular pre-determined object shapes and sizes. If we expect future robots to be more present in our factories and our homes, they must be able to adapt to new environments and objects and learn new tasks without being constantly reprogrammed by expert users. Noticing the limitations of current manipulation systems and the great advancements in machine learning over the last decade, it is natural to ask how learning methods can help create more capable and adaptive robots. In this thesis, we ask how learning methods can be used to improve the performance of the robotic manipulation pipeline in the subproblems of motion planning, grasping, and insertion.
In three separate chapters, we explore different learning approaches to these subproblems.
First, we note that the performance of optimization-based robot motion planning algorithms highly depends on initial samples, commonly obtained by running a sample-based planning method, which can be slow in high-dimensional and complex scenes and produce non-smooth solutions. Therefore, given previously solved path-planning problems, it is highly desirable to learn this data distribution and use it as a prior for new similar problems. However, this data is highly multimodal, as there are many ways to traverse an environment from a start to a goal configuration. To encode this prior knowledge and adapt to new tasks, we propose Motion Planning Diffusion (MPD), a method to learn trajectory priors for motion planning with diffusion models that, given a new cost function, adapts to new scenes and generates diverse and smooth solutions.
Second, we explore the problem of grasping objects given a single-view camera. The problem of stable grasping is encountered in almost every manipulation task, as the first step in manipulating an object is to grasp it. Particularly in real-world applications, there is often only partial information about the object due to occlusions or sensor limitations. As obtaining real-world grasp data is costly and time-consuming, simulation methods have been developed to collect large amounts of data pairs of objects and stable grasps. The generated data is highly multimodal, as there are many ways to grasp an object. To overcome these challenges, we propose Grasp Diffusion Network (GDN), a method that learns a multimodal grasp generative model given partial views of an object, by modelling the distribution with diffusion models in the manifold of homogenous transformations.
Third, we combine imitation and reinforcement learning to learn control policies for assembly tasks in environments that are difficult to model, such as contact-rich insertion tasks. To solve these problems in the real world, it is helpful to first provide a small set of demonstrations and learn a trajectory distribution with imitation learning. However, sampling a trajectory from the learned distribution and replaying the movement might not solve the task for several reasons. On the one hand, the sampled trajectory might be slightly off. On the other hand, in contact-rich tasks, a low-gain controller is commonly used, and it might not follow the trajectory precisely. Therefore, given a handful of trajectory demonstrations from a human performing an insertion task, we propose to learn Probabilistic Movement Primitives and later adapt sampled trajectories with Residual Reinforcement Learning to account for an imprecise controller and corrections in position and orientation during task execution. The proposed method is evaluated in a real-world block insertion task using a 7-dof robot that the nominal policy previously failed to solve.
In summary, this thesis contributes to advancing the field of robotic manipulation by proposing learning methods that handle large amounts of multimodal data, model the distribution in the correct manifold, and improve sample efficiency when learning from trial-and-error in contact-rich tasks. For each approach, we provide experiments in simulation and real-world tasks that show our methods can improve the performance of different components of the manipulation stack, from motion planning, grasping, and assembly tasks, and thus are a step forward in building more capable and adaptive learning-based robots
Computational Molecular Physics of FKBPs and FRB-domain containing proteins
No full textProtein-protein interactions (PPIs) are fine-tuned dynamics essential to cells and their life cycles. In cancer,
deviations of these dynamics occur, leading to uncontrollable cell growth. The process of DNA repair is
particularly important here to preserve the genetic information inside our cells and to correct mutations that can lead to changes in cell dynamics.
DNA repair is a biological pathway often controlled by large proteins that recruit other proteins. This recruitment leads to a complex construct of molecules (protein complex) that performs the actual DNA repair.
The areas within the protein complex where proteins interact with each other are called protein-protein interfaces. Understanding these interfaces at the molecular level may enable us to develop new inhibitors that interfere with DNA repair processes in a controlled manner. Since it is assumed that increased resistance to radiation and chemotherapy can be explained by the increased DNA repair ability of tumors, such inhibitors would enable the treatment of previously resistant tumors.
For the prediction of the structure of protein complexes, both classical algorithms (docking) and state-of-the-art machine learning-based methods were used in this work. The generated complexes are used as input for molecular dynamics simulations (MD simulations) to gain a time-resolved insight into the molecular biophysics of these protein-protein interfaces.
The central recruiting proteins of DNA repair form a class of protein kinases that all contain a so-called FKBP Rapamycin Binding region (FRB region). In this context, FKBP stands for FK506 binding protein, a class of proteins defined by the binding of the small molecule FK506. The FRB region is a structural subunit of a protein to which FKBPs or Rapamycin binds.
The common feature of FKBPs is the existence of a structural region binding FK506. This region can also bind numerous other proteins instead of FK506. FKBPs are therefore associated with a variety of diseases and various inhibitors interfering with these proteins’ binding capabilities are being investigated as potential drugs.
This common feature is also one of the reasons why FKBPs can fulfill various functions in our body.
This thesis examines a series of protein complexes, each containing either a single FKBP or a protein with an FRB region. Additionally, complexes in which binding occurs between both proteins are also analyzed.
The central protein kinase of the non-homologous end joining (NHEJ) DNA repair pathway was chosen as a representative for a protein containing an FRB region without a linked FKBP: DNA-dependent protein kinase, catalytic subunit (DNA-PKcs). A series of protein complexes with DNA-PKcs were modeled in collaboration with experimenters. The modeling indicates the existence of a previously unknown protein complex consisting of, among others, two DNA-PKcs molecules, which explains previously published laboratory results.
Furthermore, the protein structure of the DNA-PKcs was resolved with AlphaFold v2 (AF2) for the first time.
Based on the resolved structure, the biophysics of the binding between DNA and DNA-PKcs were investigated in the context of the NHEJ repair mechanism.
Similar studies were carried out for an FKBP of L. pneumophila. This protein is associated with the virulence of the bacterium. Several experimentally proven protein complexes containing this FKBP were modeled. This fundamental research in collaboration with the Steinert group at the TU Braunschweig provides new insight
into how the virulence of L. pneumophila can be explained
Adaptive Anomaly Detection in Network Flows with Low-Rank Tensor Decompositions and Deep Unrolling
No full textAnomaly detection (AD) is increasingly recognized as a key component for ensuring the resilience of future communication systems. While deep learning has shown state-of-the-art AD performance, its application in critical systems is hindered by concerns regarding training data efficiency, domain adaptation and interpretability. This work considers AD in network flows using incomplete measurements, leveraging a robust tensor decomposition approach and deep unrolling techniques to address these challenges. We first propose a novel block-successive convex approximation algorithm based on a regularized model-fitting objective where the normal flows are modeled as low-rank tensors and anomalies as sparse. An augmentation of the objective is introduced to decrease the computational cost. We apply deep unrolling to derive a novel deep network architecture based on our proposed algorithm, treating the regularization parameters as learnable weights. Inspired by Bayesian approaches, we extend the model architecture to perform online adaptation to per-flow and per-time-step statistics, improving AD performance while maintaining a low parameter count and preserving the problem's permutation equivariances. To optimize the deep network weights for detection performance, we employ a homotopy optimization approach based on an efficient approximation of the area under the receiver operating characteristic curve. Extensive experiments on synthetic and real-world data demonstrate that our proposed deep network architecture exhibits a high training data efficiency, outperforms reference methods, and adapts seamlessly to varying network topologies
Wavelet Enhanced LSTM for Accurate Streamflow Modeling Addressing Nonlinearity and Nonstationarity in Hydrological Data
No full textFlood prediction is challenging due to increasing climate variability and nonlinear, nonstationary hydrological data. We propose a novel hybrid model, Wavelet Long Short-Term Memory (WLSTM), combining wavelet transforms with LSTM networks. Wavelet decomposition allows separation of frequency components, enhancing LSTM’s ability to learn both short- and long-term dependencies. Applied to streamflow data from 16 stations in Hessen (2000–2017), WLSTM reduced RMSE by 66.43% and MAPE by 45.49%, while increasing R² by 2.06%. These improvements highlight its effectiveness in modeling complex hydrological dynamics. WLSTM offers a robust tool for adaptive flood risk management under climate change
Optimization Technique for Electric Bus Fleet Schedules with a-Priori Trip Disaggregation
No full textElectric vehicles represent a great opportunity to electrify the public transportation sector, and thereby mitigate greenhouse gas emissions and noise pollution among others. Due to the limited range of electric buses compared to their internal combustion engine counterparts, the scheduling process has to be rethought. A fleet operator has to decide between different charging strategies such as depot-charging, opportunity-charging, or a combination of both. The existence of different vehicle types, e. g., regarding battery capacity and passenger capacity, further complicates the matter. We introduce a multiple vehicle type scheduling problem with partial charging (MVTSP-PC) that can optimize existing fleet schedules. Our approach is based on the disaggregation of the existing bus schedule. Then, with an exact solution approach using mixed-integer linear programming and route restriction strategies we identify optimal schedules while considering computational resources. With the presented approach bus schedules for whole weeks can be efficiently re-adapted. The approach designs the optimal schedule on a daily basis, and subsequently uses a heuristic approach for the vehicle-to-route assignment when concatenating the days