#15 · ML / AI methods

Pretrained and conditioned models guide DNA, RNA, protein, and antibody design

The common thread is conditioning biomolecular models on explicit variables, with pretrained or learned priors supplying biological structure. LOCOGEN reformulates diffusion guidance for regulatory DNA design as a locality problem, damping out-of-distribution guidance while protecting learned motif positions. iGC learns a somatic hypermutation prior from B-cell lineage trees and optimizes antibody binding through reinforcement learning with a KL constraint. LIMBER, a hybrid self-attention and Mamba RNA foundation model pretrained on 40.5M ncRNA sequences, targets RNA structure and function. Ernie-Switcher applies the RNAErnie foundation model to toehold switch design, reporting up to 35-fold ON/OFF ratios, while PRIDA couples pretrained protein and RNA language models with conditional domain-adversarial learning. What is new across these items is that the conditioning variable or constraint is made explicit, whether as a locality term, a KL penalty, or a domain condition. The items are heterogeneous: LOCOGEN, iGC, and ProteoCraft are design methods, whereas PRIDA and LIMBER address prediction and interpretation.