Publications

A deep decoder neural network is applied towards the enhancement of recent algorithms for holographic Coherent Diffraction Imaging (CDI). This method does not re- quire training data, and provides improved imaging given noisy low-photon CDI data.

Transcription factors (TFs) often function as a module including both master factors and mediators binding at cis-regulatory regions to modulate nearby gene transcription. ChIP-seq profiling of multiple TFs makes it feasible to infer functional TF modules. However, when inferring TF modules based on co-localization of ChIP-seq peaks, often many weak binding events are missed, especially for mediators, resulting in incomplete identification of modules. To address this problem, we develop a ChIP-seq data-driven Gibbs Sampler to infer Modules (ChIP-GSM) using a Bayesian framework that integrates ChIP-seq profiles of multiple TFs. ChIP-GSM samples read counts of module TFs iteratively to estimate the binding potential of a module to each region and, across all regions, estimates the module abundance. Using inferred module-region probabilistic bindings as feature units, ChIP-GSM then employs logistic regression to predict active regulatory elements. Validation of ChIP-GSM predicted regulatory regions on multiple independent datasets sharing the same context confirms the advantage of using TF modules for predicting regulatory activity. In a case study of K562 cells, we demonstrate that the ChIP-GSM inferred modules form as groups, activate gene expression at different time points, and mediate diverse functional cellular processes. Hence, ChIP-GSM infers biologically meaningful TF modules and improves the prediction accuracy of regulatory region activities.

Nascent platforms for programmable quantum simulation offer unprecedented access to new regimes of far-from-equilibrium quantum many-body dynamics in (approximately) isolated systems. Here, achieving precise control over quantum many-body entanglement is an essential task for quantum sensing and computation. Extensive theoretical work suggests that these capabilities can enable dynamical phases and critical phenomena that exhibit topologically-robust methods to create, protect, and manipulate quantum entanglement that self-correct against large classes of errors. However, to date, experimental realizations have been confined to classical (non-entangled) symmetry-breaking orders. In this work, we demonstrate an emergent dynamical symmetry protected topological phase (EDSPT), in a quasiperiodically-driven array of ten ${}^{171}Yb^{+}$ hyperfine qubits in Honeywell's System Model H1 trapped-ion quantum processor. This phase exhibits edge qubits that are dynamically protected from control errors, cross-talk, and stray fields. Crucially, this edge protection relies purely on emergent dynamical symmetries that are absolutely stable to generic coherent perturbations. This property is special to quasiperiodically driven systems: as we demonstrate, the analogous edge states of a periodically driven qubit-array are vulnerable to symmetry-breaking errors and quickly decohere. Our work paves the way for implementation of more complex dynamical topological orders that would enable error-resilient techniques to manipulate quantum information.

Experimental approaches to study tissue specificity enable insight into the nature and organization of the cell types and tissues that constitute complex multicellular organisms. Machine learning provides a powerful tool to investigate and interpret tissue-specific experimental data. In this Review, we first provide a brief introduction to key single-cell and whole-tissue approaches that allow investigation of tissue specificity and then highlight two classes of machine-learning-based methods, which can be applied to analyse, model and interpret these experimental data. Deep learning methods can predict tissue-dependent effects of individual mutations on gene expression, alternative splicing and disease phenotypes. Network-based approaches can capture relationships between biomolecules, integrate large heterogeneous data compendia to model molecular circuits and identify tissue-specific functional relationships and regulatory connections. We conclude with an outlook to future possibilities in examining multicellular complexity by combining high-resolution, large-scale multiomics data sets and interpretable machine learning models.

Modern high-throughput sequencing technologies provide low-cost microbiome survey data across all habitats of life at unprecedented scale. At the most granular level, the primary data consist of sparse counts of amplicon sequence variants or operational taxonomic units that are associated with taxonomic and phylogenetic group information. In this contribution, we leverage the hierarchical structure of amplicon data and propose a data-driven and scalable tree-guided aggregation framework to associate microbial subcompositions with response variables of interest. The excess number of zero or low count measurements at the read level forces traditional microbiome data analysis workflows to remove rare sequencing variants or group them by a fixed taxonomic rank, such as genus or phylum, or by phylogenetic similarity. By contrast, our framework, which we call trac (tree-aggregation of compositional data), learns data-adaptive taxon aggregation levels for predictive modeling, greatly reducing the need for user-defined aggregation in preprocessing while simultaneously integrating seamlessly into the compositional data analysis framework. We illustrate the versatility of our framework in the context of large-scale regression problems in human gut, soil, and marine microbial ecosystems. We posit that the inferred aggregation levels provide highly interpretable taxon groupings that can help microbiome researchers gain insights into the structure and functioning of the underlying ecosystem of interest.

High-precision photometric observations have revealed ubiquitous stochastic low-frequency photometric variability in early-type stars. It has been suggested that this variability arises due to either subsurface convection or internal gravity waves launched by the convective core. Here we show that relevant properties of convection in subsurface convective layers correlate very well with the timescale and amplitude of stochastic low-frequency photometric variability, as well as with the amplitude of macroturbulence. We suggest that low-frequency, stochastic photometric variability and surface turbulence in massive stars are caused by the presence of subsurface convection. We show that an explanation for the observed surface photometric variability and macroturbulence relying on convective core driven internal gravity waves encounters a number of difficulties and seems unlikely to be able to explain the observed trends.

A perturbed black hole rings down by emitting gravitational waves in tones with specific frequencies and durations. Such tones encode prized information about the geometry of the source spacetime and the fundamental nature of gravity, making the measurement of black hole ringdowns a key goal of gravitational wave astronomy. However, this task is plagued by technical challenges that invalidate the naive application of standard data analysis methods and complicate sensitivity projections. In this paper, we provide a comprehensive account of the formalism required to properly carry out ringdown analyses, examining in detail the foundations of recent observational results, and providing a framework for future measurements. We build on those insights to clarify the concepts of ringdown detectability and resolvability -- touching on the drawbacks of both Bayes factors and naive Fisher matrix approaches -- and find that overly pessimistic heuristics have led previous works to underestimate the role of ringdown overtones for black hole spectroscopy. We put our framework to work on the analysis of a variety of simulated signals in colored noise, including analytic injections and a numerical relativity simulation consistent with GW150914. We demonstrate that we can use tones of the quadrupolar angular harmonic to test the no-hair theorem at current sensitivity, with precision comparable to published constraints from real data. Finally, we assess the role of modeling systematics, and project measurements for future, louder signals. We release ringdown, a Python library for analyzing black hole ringdowns using the the methods discussed in this paper, under a permissive open-source license at \href{https://github.com/maxisi/ringdown}{this https URL}.

Relativistic magnetized shocks are a natural source of coherent emission, offering a plausible radiative mechanism for Fast Radio Bursts (FRBs). We present first-principles 3D simulations that provide essential information for the FRB models based on shocks: the emission efficiency, spectrum, and polarization. The simulated shock propagates in an e± plasma with magnetization σ>1. The measured fraction of shock energy converted to coherent radiation is ≃10−3σ−1, and the energy-carrying wavenumber of the wave spectrum is ≃4ωc/c, where ωc is the upstream gyrofrequency. The ratio of the O-mode and X-mode energy fluxes emitted by the shock is ≃0.4σ−1. The dominance of the X-mode at σ≫1 is particularly strong, approaching 100% in the spectral band around 2ωc. We also provide a detailed description of the emission mechanism for both X- and O-modes.

CRISPR/Cas9 is a revolutionary gene-editing technology that has been widely utilized in biology, biotechnology and medicine. CRISPR/Cas9 editing outcomes depend on local DNA sequences at the target site and are thus predictable. However, existing prediction methods are dependent on both feature and model engineering, which restricts their performance to existing knowledge about CRISPR/Cas9 editing

We characterize and remedy a failure mode that may arise from multi-scale dynamics with scale imbalances during training of deep neural networks, such as Physics Informed Neural Networks (PINNs). PINNs are popular machine-learning templates that allow for seamless integration of physical equation models with data. Their training amounts to solving an optimization problem over a weighted sum of data-fidelity and equation-fidelity objectives. Conflicts between objectives can arise from scale imbalances, heteroscedasticity in the data, stiffness of the physical equation, or from catastrophic interference during sequential training. We explain the training pathology arising from this and propose a simple yet effective inverse-Dirichlet weighting strategy to alleviate the issue. We compare with Sobolev training of neural networks, providing the baseline of analytically ϵ-optimal training. We demonstrate the effectiveness of inverse-Dirichlet weighting in various applications, including a multi-scale model of active turbulence, where we show orders of magnitude improvement in accuracy and convergence over conventional PINN training. For inverse modeling using sequential training, we find that inverse-Dirichlet weighting protects a PINN against catastrophic forgetting.