Improving global school connectivity is critical for ensuring inclusive and equitable quality education. To reliably estimate the cost of connecting schools, governments and connectivity providers require complete and accurate school location data – a resource that is often scarce in many low- and middle-income countries. To address this challenge, we propose a cost-effective, scalable approach to locating schools in high-resolution satellite images using weakly supervised deep learning techniques. Our best models, which combine vision transformers and convolutional neural networks, achieve AUPRC values above 0.96 across 10 pilot African countries. Leveraging explainable AI techniques, our approach can approximate the precise geographical coordinates of the school locations using only low-cost, classification-level annotations. To demonstrate the scalability of our method, we generate nationwide maps of school location predictions in African countries and present a detailed analysis of our results, using Senegal as our case study. Finally, we demonstrate the immediate usability of our work by introducing an interactive web mapping tool to streamline human-in-the-loop model validation efforts by government partners. This work successfully showcases the real-world utility of deep learning and satellite images for planning regional infrastructure and accelerating universal school connectivity.

RNA-protein binding is critical to gene regulation, controlling fundamental processes including splicing, translation, localization and stability, and aberrant RNA-protein interactions are known to play a role in a wide variety of diseases. However, molecular understanding of RNA-protein interactions remains limited; in particular, identification of RNA motifs that bind proteins has long been challenging, especially when such motifs depend on both sequence and structure. Moreover, although RNA binding proteins (RBPs) often contain more than one binding domain, algorithms capable of identifying more than one binding motif simultaneously have not been developed. In this paper we present a novel pipeline to determine binding peaks in crosslinking immunoprecipitation (CLIP) data, to discover multiple possible RNA sequence/structure motifs among them, and to experimentally validate such motifs. At the core is a new semi-automatic algorithm SARNAclust, the first unsupervised method to identify and deconvolve multiple sequence/structure motifs simultaneously. SARNAclust computes similarity between sequence/structure objects using a graph kernel, providing the ability to isolate the impact of specific features through the bulge graph formalism. Application of SARNAclust to synthetic data shows its capability of clustering 5 motifs at once with a V-measure value of over 0.95, while GraphClust achieves only a V-measure of 0.083 and RNAcontext cannot detect any of the motifs. When applied to existing eCLIP sets, SARNAclust finds known motifs for SLBP and HNRNPC and novel motifs for several other RBPs such as AGGF1, AKAP8L and ILF3. We demonstrate an experimental validation protocol, a targeted Bind-n-Seq-like high-throughput sequencing approach that relies on RNA inverse folding for oligo pool design, that can validate the components within the SLBP motif. Finally, we use this protocol to experimentally interrogate the SARNAclust motif predictions for protein ILF3. Our results support a newly identified partially double-stranded UUUUUGAGA motif similar to that known for the splicing factor HNRNPC.

MOTIVATION: RNA thermometers (RNATs) are cis-regulatory elements that change secondary structure upon temperature shift. Often involved in the regulation of heat shock, cold shock and virulence genes, RNATs constitute an interesting potential resource in synthetic biology, where engineered RNATs could prove to be useful tools in biosensors and conditional gene regulation. RESULTS: Solving the 2-temperature inverse folding problem is critical for RNAT engineering. Here we introduce RNAiFold2T, the first Constraint Programming (CP) and Large Neighborhood Search (LNS) algorithms to solve this problem. Benchmarking tests of RNAiFold2T against existent programs (adaptive walk and genetic algorithm) inverse folding show that our software generates two orders of magnitude more solutions, thus allowing ample exploration of the space of solutions. Subsequently, solutions can be prioritized by computing various measures, including probability of target structure in the ensemble, melting temperature, etc. Using this strategy, we rationally designed two thermosensor internal ribosome entry site (thermo-IRES) elements, whose normalized cap-independent translation efficiency is approximately 50% greater at 42 °C than 30 °C, when tested in reticulocyte lysates. Translation efficiency is lower than that of the wild-type IRES element, which on the other hand is fully resistant to temperature shift-up. This appears to be the first purely computational design of functional RNA thermoswitches, and certainly the first purely computational design of functional thermo-IRES elements. AVAILABILITY: RNAiFold2T is publicly available as part of the new release RNAiFold3.0 at https://github.com/clotelab/RNAiFold and http://bioinformatics.bc.edu/clotelab/RNAiFold, which latter has a web server as well. The software is written in C ++ and uses OR-Tools CP search engine. CONTACT: [email protected] SUPPLEMENTARY INFORMATION: Supplementary data are available at Bioinformatics online.

Abstract Background RNA inverse folding is the problem of finding one or more sequences that fold into a user-specified target structure s0, i.e. whose minimum free energy secondary structure is identical to the target s0. Here we consider the ensemble of all RNA sequences that have low free energy with respect to a given target s0. Results We introduce the program , which computes the dual partition functionZ∗, defined as the sum of Boltzmann factors exp(−E(a,s0)/RT) of all RNA nucleotide sequences a compatible with target structure s0. Using , we efficiently sample RNA sequences that approximately fold into s0, where additionally the user can specify IUPAC sequence constraints at certain positions, and whether to include dangles (energy terms for stacked, single-stranded nucleotides). Moreover, since we also compute the dual partition functionZ∗(k) over all sequences having GC-content k, the user can require that all sampled sequences have a precise, specified GC-content. Using Z∗, we compute the dual expected energy 〈E∗〉, and use it to show that natural RNAs from the 12.0 database have higher minimum free energy than expected, thus suggesting that functional RNAs are under evolutionary pressure to be only marginally thermodynamically stable. We show that C. elegans precursor microRNA (pre-miRNA) is significantly non-robust with respect to mutations, by comparing the robustness of each wild type pre-miRNA sequence with 2000 [resp. 500] sequences of the same GC-content generated by , which approximately [resp. exactly] fold into the wild type target structure. We confirm and strengthen earlier findings that precursor microRNAs and bacterial small noncoding RNAs display plasticity, a measure of structural diversity. Conclusion We describe , which rapidly computes the dual partition functionZ∗ and samples sequences having low energy with respect to a target structure, allowing sequence constraints and specified GC-content. Using different inverse folding software, another group had earlier shown that pre-miRNA is mutationally robust, even controlling for compositional bias. Our opposite conclusion suggests a cautionary note that computationally based insights into molecular evolution may heavily depend on the software used. C/C++-software for is available at http://bioinformatics.bc.edu/clotelab/RNAdualPF .

Epithelial morphogenesis generates the shape of tissues, organs and embryos and is fundamental for their proper function. It is a dynamic process that occurs at multiple spatial scales from macromolecular dynamics, to cell deformations, mitosis and apoptosis, to coordinated cell rearrangements that lead to global changes of tissue shape. Using time lapse imaging, it is possible to observe these events at a system level. However, to investigate morphogenetic events it is necessary to develop computational tools to extract quantitative information from the time lapse data. Toward this goal, we developed an image-based computational pipeline to preprocess, segment and track epithelial cells in 4D confocal microscopy data. The computational pipeline we developed, for the first time, detects the adherens junctions of epithelial cells in 3D, without the need to first detect cell nuclei. We accentuate and detect cell outlines in a series of steps, symbolically describe the cells and their connectivity, and employ this information to track the cells. We validated the performance of the pipeline for its ability to detect vertices and cell-cell contacts, track cells, and identify mitosis and apoptosis in surface epithelia of Drosophila imaginal discs. We demonstrate the utility of the pipeline to extract key quantitative features of cell behavior with which to elucidate the dynamics and biomechanical control of epithelial tissue morphogenesis. We have made our methods and data available as an open-source multiplatform software tool called TTT (http://github.com/morganrcu/TTT).

Protein structure prediction is regarded as a highly challenging problem both for the biology and for the computational communities. In recent years, many approaches have been developed, moving to increasingly complex lattice models and off-lattice models. This paper presents a Large Neighborhood Search (LNS) to find the native state for the Hydrophobic-Polar (HP) model on the Face-Centered Cubic (FCC) lattice or, in other words, a self-avoiding walk on the FCC lattice having a maximum number of H-H contacts. The algorithm starts with a tabu-search algorithm, whose solution is then improved by a combination of constraint programming and LNS. The flexible framework of this hybrid algorithm allows an adaptation to the Miyazawa-Jernigan contact potential, in place of the HP model, thus suggesting its potential for tertiary structure prediction. Benchmarking statistics are given for our method against the hydrophobic core threading program HPstruct, an exact method which can be viewed as complementary to our method.

There are several evolutionary approaches for solving random binary Constraint Satisfaction Problems (CSPs). In most of these strategies we find a complex use of information regarding the problem at hand. Here we present a hybrid Evolutionary Algorithm that outperforms previous approaches in terms of effectiveness and compares well in terms ofefficiency. Our algorithm is conceptual and simple, featuring a GRASP-like (GRASP stands for Greedy Randomized Adaptive Search Procedure) mechanism for genotype-to-phenotype mapping, and without considering any specific knowledge of the problem. Therefore, we provide a simple algorithm that harnesses generality while boosting performance.

Finding Golomb rulers is an extremely challenging optimization problem (with many practical applications) that has been approached by a variety of search methods in recent years. This paper presents a hybrid evolutionary algorithm to find near-optimal Golomb rulers in reasonable time. The algorithm, which is conceptual simple and uses a natural modeling, focuses on feasibility, finding near-optimal rulers indirectly. It significantly outperforms earlier (hybrid) evolutionary algorithms and compares favorably with hybridizations of local search and constraint programming. In particular, the algorithm quickly finds optimal rulers with up to 11 marks and isolates optimal rulers with up to 14 marks in reasonable time. It also finds near-optimal rulers for up to 16 marks quickly.