Protein *sequence* design has seen significant advances through discrete diffusion and autoregressive approaches, yet the potential of continuous diffusion remains underexplored. Here, we present *DiMA*, a latent diffusion framework that operates on protein language model representations. Through systematic exploration of architectural choices and diffusion components, we develop a robust methodology that generalizes across multiple protein encoders ranging from 8M to 3B parameters. We demonstrate that our framework achieves consistently high performance across sequence-only (ESM-2, ESMc), dual-decodable (CHEAP), and multimodal (SaProt) representations using the same architecture and training approach. We conduct extensive evaluation of existing methods alongside *DiMA* using multiple metrics across two protein modalities, covering quality, diversity, novelty, and distribution matching of generated proteins. *DiMA* consistently produces novel, high-quality and diverse protein sequences and achieves strong results compared to baselines such as autoregressive, discrete diffusion and flow matching language models. The model demonstrates versatile functionality, supporting conditional generation tasks including protein family-generation, motif scaffolding and infilling, and fold-specific sequence design, despite being trained solely on sequence data. This work provides a universal continuous diffusion framework for protein sequence generation, offering both architectural insights and practical applicability across various protein design scenarios. Code is released at [GitHub](https://github.com/MeshchaninovViacheslav/DiMA).

A key challenge in protein engineering is understanding how mutations affect protein fitness and stability. Most of current state-of-the-art models fine-tune protein structure prediction or protein language models or even pretrain their own. Despite its widespread use within computational workflows, AlphaFold2 exhibits limited sensitivity in assessing the effects of amino acid point mutations on protein structure, thereby constraining its utility in sequence design and protein engineering. In this work, we propose a simple modification of AlphaFold2 inference that improves the model's capacity to capture the structural impacts of amino acid mutations. We achieve this by discarding the multiple sequence alignment and masking the template in recycling stages. Moreover, we introduce AFToolkit, a framework that leverages the embeddings of the modified AlphaFold2 model and simple adapter models to solve multiple protein engineering tasks. In contrast to other methods, our approach does not require fine-tuning the AlphaFold2 model or pretraining a new model from scratch on large datasets. It also supports handling multiple mutations, insertions, and deletions by directly modifying the input protein sequence. The proposed approach achieves strong performance across established benchmarks in terms of Spearman correlation: $0.68$ on PTMul, $0.60$ on cDNA-indel, and $0.57$ on C380.

MOTIVATION: Accurate prediction of change in protein stability due to point mutations is an attractive goal that remains unachieved. Despite the high interest in this area, little consideration has been given to the transformer architecture, which is dominant in many fields of machine learning. RESULTS: In this work, we introduce PROSTATA, a predictive model built in a knowledge-transfer fashion on a new curated dataset. PROSTATA demonstrates advantage over existing solutions based on neural networks. We show that the large improvement margin is due to both the architecture of the model and the quality of the new training dataset. This work opens up opportunities to develop new lightweight and accurate models for protein stability assessment. AVAILABILITY AND IMPLEMENTATION: PROSTATA is available at https://github.com/AIRI-Institute/PROSTATA and https://prostata.airi.net.