Analyses routinely conducted in both fundamental research and biomedical studies often report aggregate signals derived from heterogeneous mixtures of cells. These mixtures reflect the intrinsic anatomical and functional complexity of the tissues and organs from which they are derived. Such bulk measurements obscure cell-specific signals, thereby hindering accurate interpretation of the underlying biological processes. Here, we introduce CLIMB and CLIFF, a two-step machine-learning framework that deconvolves bulk RNA-seq into cell-type composition and expression, and then resolves bulk drug sensitivity measurements into cell subtype- and single-cell-specific profiles. We first validate CLIMB using in-vitro mixtures of leukemic cell lines with experimentally known cell-type proportions, then extend this validation to large-scale, cross-dataset pseudo-bulk analyses built from independent scRNA-seq references across two disease contexts and five cancer types — mimicking the reference-to-target mismatches encountered in real applications — and show that CLIMB accurately infers the cellular composition of bulk samples, outperforming state-of-the-art bulk deconvolution methods. We emphasize CLIMB’s effectiveness in acute myeloid leukemia, where it offers highly accurate estimation of leukemic stem cell fractions across three test datasets and three patient cohorts. We then validate CLIFF, our algorithm for estimating cell subtype- and single-cell-specific drug sensitivity, first on the same in-vitro cell line mixtures using experimentally measured drug responses, and then through in-depth, realistic-scale pseudo-bulk simulations. We applied our deconvolution strategy to the beatAML cohort dataset, currently the most extensive database of ex-vivo drug screening data. This analysis generated a cell subtype–resolved drug sensitivity landscape for 121 compounds, revealing multiple compounds putatively active against leukemic stem cells, including midostaurin, A-674563, SNS-032, and foretinib. Collectively, our work introduces a general computational framework for resolving bulk functional measurements at cell-type and single-cell resolution, providing an attractive new computational tool for drug development and precision medicine.

Avec l’application d’une structure ponctuelle dans un immeuble de logement coopératif à Genève, le bureau Dreier Frenzel a exploré, entre 2010 et 2018, les possibilités d’une architecture non-conformiste : la juxtaposition d’appartements pour une famille nucléaire traditionnelle et de clusters comptant jusqu’à 25 pièces.

À Genève, la forme des quartiers dont on hérite et de ceux qu’on continue à construire est déterminée par des outils légaux, réglementaires et financiers élaborés dans les années 1950 pour répondre à la pénurie de logements. Un retour sur ces différents outils et sur quelques projets emblématiques qui en sont le fruit permet d’éclairer la production contemporaine de «nouveaux quartiers».

Der Ausbau der Wasserkraft in der Schweiz erfolgte dank der Arbeit tausender Saisonniers, die teils ihr Leben auf den Baustellen der Stauseen liessen. L'histoire de l'électricité en Suisse est donc aussi une histoire de migrations. Ein Beitrag in Deutsch (erster Teil) und Französisch (zweiter Teil) fruit d'une collaboration entre deux auteurs issus du domaine.

Accurate predictions with well-calibrated uncertainties are vital for safety-critical LiDAR scene understanding in autonomous driving and robotics. While Bayesian Neural Networks and deep ensembles are effective, they are computationally expensive, requiring multiple forward passes or independently trained models. We propose GMM-NIG, a Bayesian generative classification head that extends the GMMSeg framework by placing conjugate Normal-Inverse-Gamma (NIG) priors over class-conditional GMM parameters to obtain posterior-predictive uncertainty. While the backbone remains deterministic, GMM-NIG enables analytic closed-form posterior updates in a single training run. In inference, GMM-NIG yields a mixture of Student-t distributions, providing well-calibrated predictive uncertainty without extra computational overhead. On SemanticKITTI and nuScenes, GMM-NIG achieves state-of-the-art calibration with an Adaptive Calibration Error of 2.01%, outperforming deep ensembles, Monte Carlo Dropout, evidential models and post-hoc calibration methods. Beyond calibration, it increases segmentation mIoU by nearly 6% and yields the strongest failure detection performance.

Coral reefs are among the most biodiverse and most threatened ecosystems on Earth, declining under ocean warming and recurrent mass bleaching. Whether they persist depends on how fast corals can adapt. In this context, this thesis assesses the adaptive potential of the reef-building coral Acropora millepora in the Coral Sea from a holobiont perspective, reading the environmental seascape, the host genome, the algal symbionts and the colony phenotype together across fourteen reef sites in two contrasting regions: the Capricorn Bunker Group, at the southern tip of the Australian Great Barrier Reef, and the remote Chesterfield-Bellona plateau of New Caledonia.

The environmental seascape was characterised from satellite remote sensing, as eighty descriptors of thermal stress, water quality, salinity and currents computed over four time windows, and was used to guide sampling across the environmental range rather than across geography alone. Colony morphology, colour, health and microhabitat were scored from standardised in situ photographs of 268 colonies. The host genome of 236 colonies was resolved by low-coverage whole genome sequencing (lcWGS), and the algal symbionts of 237 colonies by ITS2 metabarcoding. Genome-environment associations were then scanned with LFMM2 and SAMbada, cross-checked against the environment-free PCAdapt selection scan, and integrated, alongside the symbionts and the phenotype, in redundancy analyses that projected a polygenic adaptive index across the reefs.

The two regions form two genetically distinct units, each effectively panmictic within itself, separated by low but consistent differentiation and with Chesterfield-Bellona carrying about three times the nucleotide diversity of the Capricorn Bunker Group. Within Chesterfield-Bellona, genetic differentiation increases with geographic distance, a stepping-stone pattern absent from the Capricorn. Phenotype, symbiont community and host genome separate the same two regions, the dominant Cladocopium type profile switching between them. Candidate loci associated with sea surface temperature anomalies and heat-stress metrics fall in molecular chaperones, an antioxidant transporter, and extracellular-matrix and cilia genes, a set that recurs in independent global coral studies. The adaptive index draws a warm, clear-water-adapted pole across the Chesterfield-Bellona plateau, strongest in its north, grading to a cooler, more turbid-adapted pole in the Capricorn Bunker Group. It is on the Chesterfield-Bellona plateau that the adaptive signal of the host, the symbiont community and colony health agree, a first step towards seascape genomics at the scale of the holobiont.

Because the two regions lie far apart, the regional contrast between them is confounded with the environment, the symbionts and the host structure, so adaptive signals within each region were harder to resolve. However, sampled in June and September 2023, months before the severe 2024 El Niño marine heatwave, the colonies offer a rare opportunity to test in the field whether the genotypes predicted to be heat-adapted actually survived. Conserving the adaptive potential of corals, in a climate that no longer holds still, means reading the host, the symbiont and the environment together.

Quantitative chemical analysis in the scanning transmission electron microscope (STEM) increasingly relies on machine learning to extract phase distributions and compositions from energy-dispersive X-ray spectroscopy (EDXS) data. For example, the physics-guided factorisation method ESpM-NMF, which constrains non-negative matrix factorisation with an explicit model of X-ray emission, delivers interpretable and directly quantified results. However, its implementation resides on certain assumptions and limitations: X-ray absorption is corrected using a single thickness value and a single density value for the entire dataset; spatially-sparse minority phases are lost below the statistical noise floor; and monolithic processing of large datasets demands prohibitive amounts of memory. The present thesis addresses the first of these limitations by innovating a physics-based, spatially-dependent absorption correction using information from co-acquired electron energy-loss spectroscopy (EELS) datasets. The next two challenges are then addressed by processing the data in blocks, without loading the entire dataset into memory.

An extra component of the thesis is the creation of experimental test samples for algorithm validation, going beyond the common approach of validation using entirely synthetic datasets. To this end, two model systems with lithographically-defined ground truth were nanofabricated. The first buries patterned Al lines at controlled depths within a Ge matrix, so that a single lamella samples a systematic variation of composition and absorption path. The second carries sparse Ag structures on electron-transparent Si3N4 membranes, occupying only ~0.15% of the field of view, thereby emulating an extreme minority-phase scenario under fully known conditions.

A spatial registration framework is developed to bring independently-acquired datasets into pixel-wise correspondence, combining an affine transformation with a mutual information metric and an evolutionary optimiser. Building on this, an EELS-informed absorption correction is devised: composition estimates from an initial ESpM-NMF decomposition are combined with low-loss EELS thickness maps to compute a pixel-wise, energy-dependent correction, which converges in a single iteration. Tested on perovskite oxide references and on the nanofabricated buried structures, it obtains stoichiometries closer to the expected values, improving light-element quantification by up to ten atomic percentage points and proving inherently robust to thickness and porosity variations.

To overcome memory and minority-phase limitations, the block-wise decomposition scheme is developed. Statistical decompositions discard chemically-distinct minor phases whose contribution to total variance is too small. Spatial partitioning increases their local statistical weight; locally-harvested spectral signatures are then classified into global phases and refined into seamless full-field maps. Spatial partitioning recovers phases invisible to monolithic decompositions, resolving the complete Ag-L line family from structures covering a fraction of a percent of the field of view, and reduces peak memory usage by an order of magnitude.

All tools are distributed as ASTORUM, an open-source package written mainly in C++ with a Python interface for Jupyter-notebook integration, extending the quantitative, physics-guided analysis framework of ESpM-NMF to dataset sizes and phase fractions previously out of reach.

Flavour-changing neutral current (FCNC) transitions such as b -> s mu+ mu- are forbidden at tree level in the Standard Model and proceed only through loop diagrams. Their consequently low rates make them sensitive probes of physics beyond the Standard Model at mass scales well above those directly accessible at colliders. This thesis presents a measurement of the differential branching fraction of the FCNC decay B+ -> K+ pi+ pi- mu+ mu-, in bins of the invariant mass squared of the dimuon system, q^2. The measurement uses proton-proton collision data recorded by the LHCb experiment during Run 1 and Run 2 of the LHC, corresponding to an integrated luminosity of 9 fb^-1. Signal candidates are isolated using a combination of cut-based requirements, dedicated vetoes against misidentified backgrounds, and boosted decision tree classifiers. Efficiencies are determined from simulation, corrected using data-driven techniques derived from the B+ -> K+ pi+ pi- J/psi(-> mu+ mu-) control mode, and modelled as a function of the decay phase space using a machine-learning approach that accommodates the high dimensionality of the five-body final state. Signal and normalisation yields are extracted from fits to the B+ invariant-mass distribution, and the branching fraction is measured relative to that of the B+ -> psi(2S)(-> mu+ mu-) K+ decay. The results present the most precise measurement of this decay channel to date, and are reported in intervals of q^2 between 0.1 and 19.0 GeV^2/c^4, including two alternative definitions of the region close to the kinematic limit. In this high-q^2 region, the signal significances are measured to be 2.8 sigma in q^2 in [15, 19] GeV^2/c^4, and 5.1 sigma in the wider q^2 in [14.3, 19] GeV^2/c^4 bin. The branching fraction results are found to agree with the previous LHCb measurement of this decay. As a cross-check, the branching fraction of the resonant B+ -> K+ pi+ pi- J/psi decay is also measured, improving substantially on the precision of the current world average. For the rare B+ -> K+ pi+ pi- mu+ mu- decay, the experimental precision remains limited by the size of the dataset, which motivates a future measurement with the recently completed LHCb Run 3 dataset; preliminary work in this direction is also discussed at the end of this thesis and shows promising results.

AI control is a field of study intended to provide a second layer of defense should our efforts at AI alignment fail. Starting with the premise that an AI system is misaligned, AI control is concerned with designing mechanisms to either prevent misbehavior or catch the AI red-handed. In this thesis, we present an overview of AI control alongside several contributions to the field, split into the following chapters:

Overview of AI control. To motivate AI control, we begin with a survey of theoretical and empirical evidence of misalignment in modern AI systems. Then, we discuss the high-level principles that guide AI control mitigations, covering common deployment contexts, the specific threats that AI control aims to prevent, and how AI control evaluations are conducted. We also introduce a distinction between concentrated and diffuse threats from untrusted AI models and discuss why these two types of threats demand different mitigations. Concentrated threats are those that unfold over a short period of time. The science of concentrated threats, sometimes called concentrated AI control, is more developed; following prior work, we present a mathematical model that can be used to evaluate the safety of concentrated AI control protocols.

Control tax. In this chapter, we study a concentrated AI control protocol called trusted monitoring. Trusted monitoring proceeds by having a weaker trusted AI model assign suspiciousness scores to the actions of the untrusted model, with the most suspicious actions escalated for human auditing. To evaluate trusted monitoring, we have language models generate solutions to competitive programming problems that contain subtle bugs. We then test how well other models can distinguish these "backdoored" solutions from the correct ones. Using the mathematical model from the previous chapter, we show how to compute the safety of trusted monitoring analytically in the limit of a large number of deployed AI agents. Finally, we combine our applied and theoretical results under the framework of the control tax---the operational and financial cost of integrating trusted monitoring into AI pipelines.

Adaptive attacks on AI control protocols. In this chapter, we present simple attacks that subvert existing monitoring-based protocols in concentrated AI control. The most common kind of trusted monitor used in control protocols is a prompted language model, and language models are notoriously susceptible to prompt injections. We exploit this vulnerability to demonstrate that trusted monitors can be induced to label most of the code they see as not suspicious, rendering them essentially useless. Our language model-generated prompt injections generalize across several monitors. Using the mathematical model from the previous chapters, we show the disastrous downstream effects of these prompt injections on the safety of multiple control protocols.

Diffuse AI control on fuzzy tasks. In this chapter, we propose a red-teaming framework for evaluating training interventions against diffuse threats from a misaligned AI, such as deliberately underperforming on difficult-to-grade tasks. We show that a misaligned AI can subvert the process of generating ML experiment ideas even when optimized against scores assigned by a weaker model. We also show that there exist prompts for the weak scorer that significantly improve its robustness, though efficiently discovering such prompts with realistic affordances remains an open problem.

A glance is often enough. From a few fleeting cues, the tilt of a shoulder, a shift in tone, a wandering gaze, we read what another person is doing and how they feel, almost without effort. This fluency is hard to reproduce in machines: the signal is sparse, noisy, and rarely labelled, and meaning rarely lives in any single observation. It emerges from relationships, across the body, between sensory signals, and over time. This thesis asks how such context should shape what a model represents about a person, developing representations that capture four progressively richer forms of it: spatiotemporal, latent, cross-modal, and historical.

We begin with robustness. Transformer-based models for skeletal action recognition tend to fixate on a handful of joints, mistaking the quirks of their training source for the structure of movement and faltering under perturbation or a change of pose estimator. By masking the joints a model relies on most and contrasting masked against standard views, MaskCLR learns coordinated body motion rather than brittle local cues. We then set supervision and reconstruction aside for prediction in latent space: S-JEPA forecasts the representations of masked joints rather than their noisy coordinates, directing capacity toward high-level action structure and yielding transferable skeletal features without labels.

We next carry latent prediction into the multimodal setting, where choices that work in isolation no longer hold. OSKAR fuses video, skeleton, and language as shared context yet predicts each modality separately, exchanging information without erasing what is specific to each, and holds its token budget fixed as modalities accumulate. Trained without annotation, it matches specialised architectures across recognition, localization, retrieval, and question answering while remaining efficient in data, parameters, and labels.

Finally, we leave the trimmed clip for the continuous first-person stream, introducing egocentric self-emotion tracking, where perception must integrate voice, surroundings, and gaze as they unfold and read a state that shifts from moment to moment. We contribute a large bilingual dataset and benchmark, and OSIRIS, a model that conditions each prediction on a few recent emotion states, capturing how affect carries over to produce coherent timelines.

Together these contributions trace a single arc, from recognizing what a person visibly does to inferring how they feel. Across it, one principle holds: human-centric perception depends less on accumulating observations than on preserving the relationships through which they acquire meaning, across space, modality, and time.