Generative modeling of intrinsically disordered protein regions by reinforcing sparse autoencoder features
IDiom is trained specifically on predicted disordered protein regions, then steered with sparse-autoencoder rewards to hit chosen functional sequence features.
The paper says IDiom-DB contains 54 million predicted IDRs curated from AlphaFold Database. Its model generates sequences matching natural IDR composition, patterning, motifs, and predicted disorder. The RL-SAE method targeted 30 features across eight design tasks and activated 90% on average, versus 24% for activation steering. The authors report better predicted localization and transcriptional activity than steering or supervised fine-tuning, plus combinations of distinct biological features in single sequences. ArXiv · AI/CL/LG's note
The paper says IDiom-DB contains 54 million predicted IDRs curated from AlphaFold Database. Its model generates sequences matching natural IDR composition, patterning, motifs, and predicted disorder. The RL-SAE method targeted 30 features across eight design tasks and activated 90% on average, versus 24% for activation steering. The authors report better predicted localization and transcriptional activity than steering or supervised fine-tuning, plus combinations of distinct biological features in single sequences. ArXiv · AI/CL/LG's note
score 5