Jason Liu

Protein design has been dominated by diffusions due to a "structure-first" perspective. What about intrinsically disordered proteins? We scale language-based design using the modern RL stack and our model IDiom. Paper: biorxiv.org/content/10.648… Try it: idiom-designer.vercel.app

Paper: doi.org/10.64898/2026.… Code: github.com/rotskoff-group… Dataset: huggingface.co/datasets/jxliu…

We show that IDiom generates IDRs that recapitulate the sequence features of natural disordered regions, and we additionally use reinforcement learning to steer generation towards sequences with compartment specific localization features.

Intrinsically disordered protein regions have remained largely out of reach for computational design. We curate 37M IDR sequences from the AlphaFold Database and train IDiom, a 122M parameter autoregressive model, as a general platform for intrinsically disordered protein design.

Generative design of intrinsically disordered protein regions with IDiom 1. The paper introduces IDiom, a 122M-parameter autoregressive (decoder-only) protein language model trained specifically for intrinsically disordered regions (IDRs), aiming to make rational design possible in a regime where structure-based generative methods do not apply. 2. Key technical idea: fill-in-the-middle training for proteins with explicit tokens that separate N-terminal context, C-terminal context, and the IDR span (, , ). This enables conditional generation of an IDR that fits into a chosen structured protein context, not just unconditional sampling. 3. Training data scale: 37 million IDRs curated from AlphaFold DB v4 using low pLDDT as a disorder proxy (plus filtering and clustering at 90% identity). They augment to 74 million sequences by also creating “context-deleted” records to train unprompted generation of fully disordered proteins (IDPs). 4. Generated sequences are diverse yet IDR-like: maximum identity to the training IDR set broadly peaks around ~60% (not memorized), length distributions match natural IDRs (mostly <100 aa with a tail to ~300 aa), and amino-acid composition recapitulates known disorder biases (e.g., enriched Pro/Ser; depleted bulky hydrophobics and aromatics vs folded CATH domains). 5. Disorder is maintained by structure prediction checks: ColabFold/AlphaFold pLDDT distributions for generated sequences closely resemble curated AFDB IDRs and experimentally validated DisProt IDRs, both for standalone IDPs and for generated IDRs evaluated within full-protein context. 6. IDiom learns “IDR grammar”, not just composition: generations reproduce natural distributions of (i) fraction of charged residues (FCR), (ii) charge patterning/blockiness (κ), (iii) hydropathy patterning (SHD), and (iv) low complexity (SEG). These metrics separate generated IDRs from folded CATH domains and align them with DisProt statistics. 7. Conditioning matters: DisProt-context–prompted generations are consistently closer to DisProt IDRs than unprompted IDPs across multiple metrics (quantified via Wasserstein-1 distances), supporting in-context learning of context-appropriate IDR features. 8. Case study (NPM1): when prompted with NPM1 flanks, IDiom generates many low-identity IDRs that still reproduce the functional charge-block architecture (κ near WT; alternating NCPR blocks), suggesting it can preserve biophysically relevant patterning without copying sequence. 9. Post-training via reinforcement learning: the authors steer IDiom with GRPO (with DAPO modification) using ProtGPS as a reward model for subcellular localization (nucleolus, chromosomes/chromatin, P-bodies, stress granules). Regularization includes KL-to-base-model, target entropy (to avoid collapse), and target length. 10. RL-induced features are biologically interpretable while staying disordered: nucleolus-targeting sequences become Lys/Arg-rich and show higher κ; chromosome-targeting sequences become Ser/Thr-rich and show strong enrichment of ELM PTM motifs; P-body and stress-granule targeting sequences enrich RNA-interaction motifs (RG/RGG, F/YGG, SYG). Importantly, generated sequences remain low-pLDDT, indicating the policy does not drift toward folded-domain priors. 💻Code: github.com/rotskoff-group… 📜Paper: biorxiv.org/content/10.648… #ComputationalBiology #ProteinDesign #IntrinsicallyDisorderedProteins #ProteinLanguageModels #Transformers #ReinforcementLearning #PhaseSeparation #SubcellularLocalization #SyntheticBiology