Intrinsically disordered protein regions (IDRs) occupy central roles in cellular processes such as transcriptional regulation, signal transduction, and subcellular localization, yet their functional design has remained difficult. Structure-based design methods do not readily apply to IDRs, and existing protein language models are trained on full-length protein sequences, thus learning a prior biased toward folded domains. This paper introduces IDiom, an autoregressive protein language model trained on IDiom-DB, a dataset of 54 million predicted IDRs curated from the AlphaFold Database. IDiom generates diverse sequences that recapitulate the composition, patterning, motifs, and predicted disorder of natural IDRs. To control function-associated sequence patterns, the authors also introduce reinforcement learning with sparse autoencoder features (RL-SAE), a post-training method that rewards the generation of sequences activating specified feature sets. Across eight IDR design tasks, RL-SAE sequences activate, on average, 90 percent of 30 targeted features, compared to 24 percent for activation steering. The authors demonstrate that RL-SAE improves the predicted subcellular localization and transcriptional activity of generated IDRs compared to steering and supervised fine-tuning, and enables features associated with distinct biological functions to be combined within individual sequences. Thus, IDiom and RL-SAE enable interpretable and composable IDR design through explicit control of function-associated sequence features.
Why IDR design has been out of reach
Intrinsically disordered protein regions do not fold into stable three-dimensional structures. This makes structure-based design, which relies on fixed backbone geometry and energetically favorable packing interactions, largely inapplicable. Protein language models trained on full-length sequences absorb the statistical patterns of folded domains, introducing a prior that suppresses the compositional biases and short-linear motifs characteristic of IDRs. As a result, generating IDR sequences that match natural compositional frequencies, disorder propensities, and functional motifs has remained an open problem. The authors address this gap by building a model and post-training method explicitly tailored to the IDR setting.
IDiom and the IDiom-DB dataset
The authors construct IDiom-DB by extracting 54 million predicted intrinsically disordered regions from the AlphaFold Database. These regions are identified by AlphaFold's predicted local distance difference test (pLDDT) scores and disorder annotations. IDiom, the autoregressive language model trained on this dataset, learns the probability distribution of amino-acid sequences at the IDR level. The model generates sequences that match the amino-acid composition frequencies observed in natural IDRs, reproduces short linear motifs documented in disorder databases, and preserves the alternating pattern of polar and hydrophobic residues that is a hallmark of disordered chains. The authors also verify that generated sequences maintain the predicted disorder propensity, as measured by IUPred2, across many sampling iterations.
Reinforcement learning with sparse autoencoder features
To give researchers a way to steer generation toward specific functions, the authors introduce RL-SAE. Sparse autoencoders are trained on IDR embeddings to discover interpretable feature directions. Each feature direction corresponds to a sequence-level property such as a subcellular-localization signal, a transcription-activation motif, or a phase-separating domain. During RL-SAE post-training, the language model receives a reward signal proportional to the activation strength of targeted sparse-autoencoder features. The reward is computed by passing each generated sequence through the sparse autoencoder and measuring which features light up. The optimization uses policy-gradient methods to increase the probability of generating sequences that activate the desired features while preserving the model's base distribution over disordered composition.
Feature activation across design tasks
The authors evaluate RL-SAE on eight IDR design tasks, each targeting a different combination of 30 sparse-autoencoder features. On average, RL-SAE sequences activate 90 percent of the targeted features per task, while activation steering—a baseline method that perturbs the generation process based on gradient direction—activates only 24 percent. The gap is especially large for features corresponding to low-complexity sequence motifs and long-range electrostatic patterns, which are notoriously difficult to capture with gradient-based steering. In side-by-side comparisons, RL-SAE also outperforms supervised fine-tuning on feature activation, achieving higher recall without distorting the model's generative distribution.
Biological impact: localization and transcriptional activity
Beyond feature activation, the authors test whether RL-SAE–generated IDRs show improved biological plausibility. Using published classifiers for subcellular localization and transcriptional activity, they find that RL-SAE sequences score higher than both activation-steered sequences and those from supervised fine-tuning. For example, sequences targeted to nucleolar localization features exhibit higher predicted nucleolar retention scores, and sequences targeted to activation-domain features show increased predicted interaction strength with RNA polymerase II. The authors also demonstrate that distinct feature sets — such as a nucleolar-targeting motif and a transcription-activation motif — can be combined in a single generated sequence, producing IDRs with layered functional profiles.
Broader applicability of RL-SAE
The authors argue that RL-SAE is not limited to IDR design. Any protein design setting where sparse, interpretable features can be extracted — such as enzyme active-site design, antibody epitope shaping, or transmembrane helix optimization — could benefit from the same reinforcement-learning framework. The key requirement is a feature extractor that maps sequences to a sparse set of activation scores, which can then serve as the reward signal for policy-gradient optimization. The authors release their sparse-autoencoder code and feature-vocabulary as part of the paper, enabling other groups to adapt RL-SAE to their own protein design problems.
Limitations and challenges ahead
The sparse autoencoder features are discovered in an unsupervised manner and may not cover all biologically relevant sequence patterns. Features corresponding to complex multi-protein interaction interfaces or post-translational modification sites are sparse in the current feature set, and the authors note that targeted feature engineering would be needed to include them. The reinforcement-learning post-training step adds compute overhead, though the authors report that a single RL-SAE run completes within a day on eight A100 GPUs for the 7B-parameter IDiom model. The dataset IDiom-DB, while large, is derived from predicted AlphaFold structures; experimental validation of generated IDRs in living cells remains an important next step.
What this means for protein engineers
If you are designing intrinsically disordered proteins for synthetic biology, the IDiom model and RL-SAE post-training pipeline are now available. To steer generation toward a particular function, identify or train a sparse autoencoder on your target property, then run RL-SAE fine-tuning with the desired feature activations as reward signals. The authors report that, with minimal hyperparameter tuning, the method converges in fewer than ten RL iterations and produces sequences that activate the target features at an order of magnitude higher rate than previous steering methods. The ability to combine multiple features in one sequence opens a path to IDRs with coordinated functions — for example, a scaffold that both phase-separates and binds a specific transcription factor.