HEIST: A Graph Foundation Model for Spatial Transcriptomics and Proteomics Data

A new methodology utilizing machine and deep learning has been introduced to effectively solve parametric integrals, demonstrating superior performance over traditional methods. This approach incorporates derivative information during training, which enhances its efficiency across various problem classes, including statistical functionals and differential equations.

NeuralOGCM has been introduced as an innovative ocean modeling framework that integrates differentiable programming with deep learning, aiming to enhance scientific simulations by balancing computational efficiency and physical fidelity. This framework features a fully differentiable dynamical solver that utilizes physics knowledge and transforms key physical parameters into learnable components, allowing for autonomous optimization through end-to-end training.

A novel method for designing high-code-rate single-IDS-correcting codewords has been introduced, leveraging deep Levenshtein distance embedding to enhance the efficiency of the 4-ary IDS channel. This development addresses the challenges posed by insertion, deletion, and substitution errors in data transmission, which have gained attention due to evolving storage and communication technologies.

A novel method called STARK has been developed to denoise spatial transcriptomics images, enhancing the identification of cell types at ultra-low sequencing depths. This technique employs adaptive regularization and kernel ridge regression, updating a spatial graph iteratively to improve image quality and gene expression interpolation.

A recent study has demonstrated that statistical physics can effectively analyze deep learning models, particularly through the lens of a multi-layer perceptron (MLP) in a supervised learning context. The research highlights the model's ability to learn rich features, especially in interpolation regimes where the number of parameters and data are closely matched.