324139 research outputs found
Sort by
Umbilical cord blood natural killer cells for adoptive immunotherapy: identifying optimal starting material and processing parameters
Get PDFIntroduction: Umbilical cord blood (UCB) is an attractive source of natural killer (NK) cells for the development of allogeneic ‘off-the-shelf’ cancer immunotherapies. This is due to the relatively high proportion of highly proliferative NK cells compared to adult peripheral blood (APB), a low risk of graft-versus-host disease and ease of procurement. However, due to the limited starting volume of UCB and naïve phenotype of isolated cells, ex vivo NK cell expansion and activation is essential to generate clinically relevant doses of cells with potent anti-tumor activity. Furthermore, intrinsic variability in both in vitro and clinical performance of NK cells from different UCB units (CBUs) has been reported. Methods: To better characterize this variability, we measured UCB NK cell ex vivo fold expansion, phenotype and cytotoxic potential using a basic expansion system. We then used these results to identify characteristics related to superior performance, enabling the optimization, selection and processing of CBUs for the manufacture of NK cells as therapies at a larger scale. Results: Our results revealed that despite wide inter-donor variability in performance between CBUs, a priori selection could be used to identify units likely to show high expansion and/or cytotoxicity. We observed that decreased time between UCB collection and CD3- UCB mononuclear cell (CBMC) isolation was associated with significantly higher NK fold expansion (n=13; p<0.05). Furthermore, a cryopreservation step following early isolation and prior to expansion, significantly increased the expansion potential of the isolated NK cells (p<0.05), thus providing an opportunity for pre-selection and parallel culture of multiple optimal units. Finally, the NK cells from CBUs collected from caesarean sections had statistically significantly increased proliferative potential compared to those from vaginal deliveries (n=13; p<0.05). Conclusion: In conclusion, early isolation and cryopreservation of CD3- CBMCs from caesarean section CBUs offer an optimal starting material for use in UCB-derived NK cell immunotherapies, providing superior ex vivo performance and enabling batch testing to selectively expand cells from CBUs with the greatest potential
Large language models miss the multi-agent mark
No full textRecent interest in Multi-Agent Systems of Large Language Models (MAS LLMs) has led to an increase in frameworks leveraging multiple LLMs to tackle complex tasks. However, much of this literature appropriates the terminology of MAS without engaging with its foundational principles. In this position paper, we highlight critical discrepancies between MAS theory and current MAS LLMs implementations, focusing on four key areas: the social aspect of agency, environment design, coordination and communication protocols, and measuring emergent behaviours. Our position is that many MAS LLMs lack multi-agent characteristics such as autonomy, social interaction, and structured environments, and often rely on oversimplified, LLM-centric architectures. The field may slow down and lose traction by revisiting problems the MAS literature has already addressed. Therefore, we systematically analyse this issue and outline associated research opportunities; we advocate for better integrating established MAS concepts and more precise terminology to avoid mischaracterisation and missed opportunities
Object localisation for scene understanding in 3D
Get PDFUnderstanding the surrounding physical world is key to the successful application of many intelligent systems, including self-driving cars and delivery robots. As a fundamental component towards this capability, object localisation has been widely studied using various sensors, particularly those operating in the optical regime such as cameras and LIDAR. The most prevalent form of object localisation is accomplished by detection and segmentation in a local sensor frame, followed by tracking over multiple frames. Thanks to its structured nature, object localisation using this approach has seen impressive progress in image-based systems. However, for complex and dynamic scenes such as roads, 3D object localisation is still not yet sufficiently accurate for reliable real-world applications. Moreover, the continually changing nature of outdoor scenes requires object localisation systems to be robust and adaptive to dynamic environments and transferable to different operating conditions. This thesis explores the challenges of 3D object localisation based on visual sensors and presents methods that improve the aforementioned issues. Firstly, to improve the performance of 3D object localisation, we study current limitations in object representations based on Cartesian coordinates, e.g., axis-aligned bounding boxes. Based on this observation, we design a Spherical Mask that uses alternative object representations based on spherical coordinates for precise 3D object localisation. Secondly, to address the issue of high cost for 3D annotation of changing real-world environments, we develop an auto-annotation method for 3D object localisation. The method utilises motion cues from the sensors as signals for finding moving objects in various road scenes. These objects serve as pseudo annotations, which can be used to train standard object localisation models. We also show that using the Spherical representation is beneficial compared to the existing box representation for finding pseudo labels. Thirdly, we explore strategies to make the best use of existing annotations and pseudo annotations acquired in the previous step. Specifically, to improve the robustness of the model in changing environments, a pseudo-label-guided clustering approach for domain adaptation is presented. Here, given the learned representation of each cluster centroid as prompt input, the localisation model's task is to continuously learn to localise objects similar to the given cluster, making the model focus more on pseudo labels, which are the centroids of the clusters. This strategy improves the robustness of the model as all existing labels are distributed to each cluster that focuses on pseudo labels. Each cluster is used by the model to find objects similar to the pseudo labels that belong to the cluster, in contrast to the existing approach that considers labels from the source environment and pseudo-labels from different environments as the same. The proposed three techniques can be combined to deploy a more precise object localisation model, which can adapt to a continuously changing environment of the real world while minimising the cost caused by human-provided annotations based on pseudo-labels found using natural cues from the sensors
A novel subgrid model for super-Eddington accretion of spinning black holes in galaxy-scale simulations
Get PDFSuper-Eddington accretion has been proposed to explain the existence of black holes (BHs) with masses exceeding a billion solar masses within the first billion years after the big bang. We present a novel accretion disc-based subgrid model for BH mass and spin evolution in the super-Eddington regime, implemented in the hydrodynamics code gizmo. In our model, motivated by results of radiation-hydrodynamics simulations of accretion discs, the growth of the BH is mediated by a subgrid accretion disc, comprising an inner photon-trapping region described by simulation-based fitting formulae and an outer thin -disc with three regions. We incorporate a self-consistent spin evolution prescription that transitions between the Bardeen–Petterson effect and inner thick-disc precession, depending on the accretion rate. We perform a suite of idealized simulations of a BH embedded in a gaseous circumnuclear disc and a spherically distributed stellar component to explore the conditions under which super-Eddington accretion can be sustained in the environment of a realistic galactic nucleus. Simulations with misaligned gas inflows onto an initially aligned BH–disc system yield very high Eddington ratios, triggered by the rapid removal of disc angular momentum via inflows. These results highlight the importance of angular momentum misalignment in enabling super-Eddington accretion and suggest that such episodes are difficult to trigger unless the system resides in a highly dynamical environment – a condition more likely to occur in high-redshift galaxies. Our model potentially provides a way to grow moderate-mass BH seeds to the sizes required to explain the bright high-redshift quasars
Between anti-racism, anti-fascism and dialogue: A reflection on practice informed by Baptist Christian tradition
No full textA presentation made at a Book launch and panel discussion on The Church, the Far Right and the Claim to Christianity, organised by the Centre for Baptist Studies and the Centre for the Study of Bible and Violence, at the Centre for Baptist Studies, Regent’s Park College, University of Oxford, 14 September 2024. A recording of other presentations made at the event can be accessed via the Centre's You Tube Channel at https://www.youtube.com/watch?v=ZfPZ7j-Uz3o, while the author's presentation can also be accessed via the Centre's You Tube channel at https://youtube. com/9kzD0N6HOO
When relics were made: vigorous stellar rotation and low dark matter content in the massive ultra-compact galaxy GS-9209 at z = 4.66
Get PDFJames Webb Space Telescope (JWST) uncovered a large number of massive quiescent galaxies (MQGs) at , which theoretical models struggle to reproduce. Explaining the number density of such objects requires extremely high conversion efficiency of baryons into stars in early dark matter haloes. Using stellar kinematics, we can investigate the processes shaping the mass assembly histories of MQGs. We present high-resolution JWST/NIRSpec integral field spectroscopy of GS-9209, a massive compact quiescent galaxy at (, pc). Full spectral fitting of the spatially resolved stellar continuum reveals a clear rotational pattern, yielding a spin parameter of . This study suggests that at least a fraction of the earliest quiescent galaxies were fast rotators and that quenching was a dynamically gentle process, preserving the stellar disc even in highly compact objects. Using Jeans anisotropic modelling and assuming a Navarro–Frenk–White profile, we measure a dark matter fraction of . Our findings use stellar kinematics to confirm the massive nature of early quiescent galaxies, previously inferred from stellar population modelling. We suggest that GS-9209 has a similar structure to low-redshift ‘relic’ galaxies. However, unlike relic galaxies, which have bottom-heavy initial mass functions (IMF), the dynamically inferred stellar mass-to-light ratio of GS-9209 is consistent with a Milky Way-like IMF. The kinematical properties of GS-9209 are different from those of early-type galaxies and more similar to those of recently quenched post-starburst galaxies at
Re-emergence of malaria in Karen State, Myanmar: has the battle of Burma been lost?
Get PDFBackground: Historically malaria has been a major cause of morbidity and mortality in the Greater Mekong Sub-region. In recent years, significant progress towards malaria elimination has been made. Myanmar harbours most of the region’s malaria burden, however after initial progress during peace time, the civil war and the COVID-19 pandemic have coincided with a resurgence of malaria. This observational study examines the resurgence of malaria in Eastern Myanmar and its contributory factors. Methods and results: Malariometric and genomic data from a long-established network of malaria clinics and village health workers in eastern Karen State serving an estimated population of 350,000 were reviewed and analysed in the context of the COVID-19 pandemic and the military coup that followed. Data from 2020 and 2024 show that the number of cases of P. falciparum malaria increased 12-fold and those of P. vivax malaria increased threefold. This resurgence was greatest in the northern parts of Karen State and coincided with reduced access to timely diagnosis and treatment. This was associated with increased malaria transmission of P. falciparum (RR = 1.72, 95% CI 1.68–1.76) and P. vivax (RR = 1.82, 95% CI 1.80–1.84). Reported malaria-related deaths remained low during the study period though underreporting cannot be excluded. Conclusion: Our study provides evidence that the disruption of services (early diagnosis and treatment) caused by the COVID-19 pandemic followed by insecurity, explains the resurgence of malaria in Karen State in Myanmar. Population movements and the clonal expansion of a specific parasite lineage were likely contributing factors. However, the decline recorded in 2024 in the number of cases and the near absence of malaria mortality showed that despite the difficulties, the malaria control system has been successful in containing the crisis. The battle for malaria control in Burma has not been lost, but the future remains uncertain
Inequalities in offline and online news media environments across six countries: the role of social class and interest in news
Get PDFIn this article, we present a comparative analysis of inequalities in online and offline news media systems across a strategic sample of six countries. Using survey data from 10 annual Digital News Report surveys (2015–2024), we compare inequalities in news use in traditional offline and online news media environments over time among internet users. Our results show that: (1) across all countries in our sample, the online news environment is more unequal than offline in terms of the number of news sources used, (2) a substantial share of the online sample does not access any news online, and (3) that education is a key driver of these inequalities. In most instances across ten annual surveys in six countries, education is a significantly stronger predictor of online news use over offline news use across all countries. However, (4) we find a mixed picture regarding the role of interest in news in fuelling inequalities. Although the ability to access news online benefits many people, these findings suggest that the ongoing shift towards more online news environments could widen the gaps between the information rich and the information poor and reinforce longstanding social inequalities in news consumption
Uterus transplants in Mexico: legal ambiguity and the case for reform
Get PDFA law amendment to Mexico’s General Health Law that would allow live uterus donation from pre-menopausal women was recently proposed. This amendment, if passed, would be the first ever piece of legislation to explicitly address uterus donation at any level of government. Two objections have been raised to the amendment, by the authorities in charge of organ transplantation. In this paper I critically examine one of the objections. The objection maintains that there is no need to reform the General Health Law, because its correct interpretation entails that live uterus donation from pre-menopausal women is already permitted. In this paper I show that the objection fails, and that uterus donation from pre-menopausal women is currently forbidden. My approach is threefold. First, I show that the objector’s interpretation of the relevant article of the General Health Law commits them to some unpalatable implications. Second, I refer to interpretative criteria set out by Mexico’s Supreme Court of Justice to show that the objector’s interpretation is flawed. Finally, I show, using an ordinary language approach, that the objector’s interpretation is flawed, because it renders the law inoperable. I conclude that a legislative amendment is justified, in order to ensure legal clarity and ethical consistency in uterine transplantation practices
FORCE: transferable visual jailbreaking attacks via Feature Over-Reliance CorrEction
Get PDFThe integration of new modalities enhances the capabilities of multimodal large language models (MLLMs) but also introduces additional vulnerabilities. In particular, simple visual jailbreaking attacks can manipulate open-source MLLMs more readily than sophisticated textual attacks. However, these underdeveloped attacks exhibit extremely limited cross-model transferability, failing to reliably identify vulnerabilities in closed-source MLLMs. In this work, we analyse the loss landscape of these jailbreaking attacks and find that the generated attacks tend to reside in highsharpness regions, whose effectiveness is highly sensitive to even minor parameter changes during transfer. To further explain the high-sharpness localisations, we analyse their feature representations in both the intermediate layers and the spectral domain, revealing an improper reliance on narrow layer representations and semantically poor frequency components. Building on this, we propose a Feature OverReliance CorrEction (FORCE) method, which guides the attack to explore broader feasible regions across layer features and rescales the influence of frequency features according to their semantic content. By eliminating non-generalizable reliance on both layer and spectral features, our method discovers flattened feasible regions for visual jailbreaking attacks, thereby improving cross-model transferability. Extensive experiments demonstrate that our approach effectively facilitates visual red-teaming evaluations against closedsource MLLMs. Our implementation is released at https: //github.com/tmllab/2026_CVPR_FORCE