IDiom: Generative Modeling of Intrinsically Disordered Protein Regions with RL-SAE Feature Control

Forum topic · 小凯 · 2026-10-04

Summary

Researchers at Stanford (Rotskoff group, with Bintu and Dunn labs) present IDiom, an autoregressive protein language model for intrinsically disordered protein regions (IDRs), trained on IDiom-DB, a dataset of 54 million predicted IDRs curated from the AlphaFold Database. IDRs govern transcriptional regulation, signal transduction, and subcellular localization, but structure-based design tools and full-length protein language models handle them poorly, since learned priors favor folded domains. IDiom generates diverse sequences that recapitulate the composition, patterning, motifs, and predicted disorder of natural IDRs. The paper also introduces RL-SAE (reinforcement learning over sparse autoencoder features), a post-training method that rewards activation of specified feature sets. Across 8 IDR design tasks, RL-SAE sequences activate on average 90% of 30 target features versus 24% for activation steering. RL-SAE outperforms steering and supervised fine-tuning in predicted subcellular localization and transcriptional activity, and can combine features tied to different biological functions within a single sequence. Code is available at https://github.com/rotskoff-group/idiom (arXiv:2610.02189).

Paper Overview

Field: Machine Learning
Authors: Jason X. Liu, Sebastian Ibarraran, Frank Hu, Soojung Yang, Xinyu A. Feng, Abigail Park, Anagha Aneesh, Lacramioara Bintu, Alexander R. Dunn, Grant M. Rotskoff
Published: 2026-10-01
arXiv: 2610.02189

Abstract

Intrinsically disordered protein regions (IDRs) play central roles in cellular processes such as transcriptional regulation, signal transduction, and subcellular localization, yet their functional design remains challenging. 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 that is biased towards folded domains.

The authors present 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, they introduce RL-SAE (reinforcement learning over sparse autoencoder features), a post-training method that rewards activation of a specified set of features.

Key Results

Across 8 IDR design tasks, RL-SAE sequences activate on average 90% of 30 target features, compared with only 24% for activation steering.
RL-SAE improves predicted subcellular localization and transcriptional activity of generated IDRs relative to both steering and supervised fine-tuning.
Features associated with different biological functions can be combined into a single generated sequence.

Significance

IDiom and RL-SAE enable interpretable, composable IDR design through explicit control over function-relevant sequence features. More broadly, RL-SAE can extend to other protein design settings where interpretable features serve as design targets.

Code: https://github.com/rotskoff-group/idiom

--- *Auto-collected on 2026-10-04*