Ibarraran, S. | bioRxiv Science

bioRxiv2026

Generative design of intrinsically disordered protein regions with IDiom

Intrinsically disordered protein regions are ubiquitous across all kingdoms of life. These structurally heterogeneous regions play central roles in cellular processes such as transcriptional regulation, cellular signaling, and subcellular organization, yet they have remained largely inaccessible to rational design. Structure-based generative methods are not applicable to proteins that lack a stable fold, and existing sequence-based approaches for disordered regions rely on sampling methods that do not capture the evolutionary statistics of natural disordered regions. Here, we introduce IDiom, an autoregressive protein language model trained on 37 million intrinsically disordered region sequences curated from the AlphaFold Database. Trained using a fill-in-the-middle data augmentation, IDiom generates disordered region sequences conditioned on their surrounding structured context, as well as fully disordered proteins without any context. The model generates diverse sequences that recapitulate biologically relevant sequence features of natural disordered regions, and we demonstrate that post-training via reinforcement learning with a subcellular localization reward model produces sequences with features which are consistent with known sequence determinants of compartment-specific localization. These results establish IDiom as a general platform for the generative design of intrinsically disordered proteins and regions.

bioinformatics↗

bioRxiv2026

Efficient, Few-shot Directed Evolution with Energy Rank Alignment

Directed evolution is a powerful and widely used technique for protein engineering, and reducing the cost of iterated experimental observations has become a major priority for practitioners. A number of recent efforts to use machine-learning-based predictors to improve sequence selection have led to remarkable improvements in efficiency, but the sparse data at each experimental iteration restricts these approaches to extremely simple models. Adapting large-scale pre-trained protein language models using experimental data offers an alternative that we show productively leverages the strong inductive biases of the natural distribution of protein sequences to navigate high-dimensional, combinatorially large fitness landscapes. Our approach uses a general-purpose "post-training" algorithm grounded in statistical physics that employs quantitative experimental rankings to directly produce a sampler for diverse, high fitness sequences with fewer data points than competing methods. The resulting adapted protein language model can itself be studied and interpreted, shedding further light on the biophysical characteristics of highly fit sequences and their properties.

bioengineering↗