Single-cell RNA sequencing (scRNA-seq) enables high-resolution exploration of cellular diversity and gene regulation, yet analyzing such data remains challenging due to technical and methodological limitations. Existing task-specific deep generative models like Variational Auto-Encoder (VAE) and its variants struggle to incorporate external biological knowledge, while transformer-based foundational large Language Models (LLMs or large LaMs) face limitations in computational cost and applicability to tabular gene expression data. Here, we introduce sciLaMA (single-cell interpretable Language Model Adapter), a novel representation learning framework that bridges these gaps by integrating static gene embeddings from multimodal LaMs with scRNA-seq tabular data through a paired-VAE architecture. Our approach generates context-aware representations for both cells and genes and outperforms state-of-the-art methods in key single-cell downstream tasks, including batch effect correction, cell clustering, and cell-state-specific gene marker and module identification, while maintaining computational efficiency. sciLaMA offers a computationally efficient, unified framework for comprehensive single-cell data analysis and biologically interpretable gene module discovery.

In this paper, we address a joint design of symbol level beam selection and precoding in the multi-user downlink beamspace multiple-input single-output (MISO) system. Unlike the general concept of eliminating or avoiding inter-user interference, the symbol level design aims to exploit the constructive interference at symbol level among the multiple users. To optimize the constructive interference while ensuring the signal interference noise ratio (SINR), a cardinality constrained minimum mean square error (MMSE) problem with per-beam power limitation is formulated. To resolve this non-convex and non-smooth problem, we propose a penalty proximal alternating linearized minimization based algorithm, in which a sequence of penalty sub-problems are solved by iteratively executing two projections: selective constructive signal region projection and sparse beam selection projection. From the extensive simulation results, the convergence of the proposed algorithm and its effectiveness are demonstrated.

Unlike the general concept of eliminating or avoiding inter-user interference in the downlink multi-user multiple-input single-output (MISO) system, the data-aided precoding scheme attempts to exploit the constructive interference at symbol level. Positive interference can be constructed to enhance the received signal gain by predicting the phase and magnitude of the inter-user interference. In this paper, we formulate a constructive interference optimization problem that minimizes a sum of minimum mean square error (MMSE) for all users, while ensuring the minimum required constructive interference gain with a fixed total power constraint. As opposed to the existing scheme, such as constructive zero-forcing precoding or minimum power precoding subject to strict phase conservation for phase-shift keying (PSK), our proposed scheme exploits a full range of relaxation for the constructive interference region, while ensuring the link performance by minimizing the sum of mean-square error (MSE) for all users. In fact, the relaxed requirements lead to more degrees of freedom for improving the cumulative constructive interference gain (CCIG) under the varying channel conditions. Furthermore, our MMSE-based optimization approach allows for a semi-closed-form optimal solution to the data-aided precoding scheme, providing a more CCIG without incurring unacceptable complexity than other state-of-the-art schemes.