DeepREAD

Adenosine deaminase acting on RNA (ADAR) converts adenosine to inosine, which the ribosome reads as guanosine. Deliver a guide RNA (gRNA) that hybridizes with a target transcript and the enzyme already in the cell will edit the resulting duplex — correcting a pathogenic mutation without touching the genome. The obstacle is that ADAR is indiscriminate: it edits whatever adenosines the duplex presents, so a guide that fixes the intended base often also edits neighboring "bystander" adenosines. The field's answer has been a handful of hand-derived heuristics — an A-C mismatch opposite the target, A-G mismatches or U-deletions across from bystanders, large internal loops — which buy specificity by giving up on-target efficiency. Machine learning helped only locally: a model trained on one target's saturation-mutagenesis screen did not transfer to another.

DeepREAD, a September 2024 bioRxiv preprint from Shape Therapeutics, attacks that transfer problem, and the name covers two modules. ActMax is an ensemble of CNNs steered by activation maximization: gradients flow back through a frozen predictor to optimize the guide sequence toward a chosen editing and specificity value. The second module is a bit diffusion model — a denoising diffusion probabilistic model adapted to discrete sequence — that treats a gRNA as a one-hot "image" in continuous space and denoises it conditioned on the target sequence and on desired ADAR1 and ADAR2 editing and specificity. Only the diffusion half generalizes beyond the training library; ActMax remains a per-target optimizer, the approach the paper casts diffusion as superseding.

#Key Features

• Target-agnostic generation: The bit diffusion model designs guides for adenosines it never saw in training, sampling from noise conditioned on a new target rather than being refit per target. It yields thousands of candidate guides in minutes, roughly 10,000 times faster than the gradient-based ActMax search.
• Conditioned on the desired outcome: Generation takes scalar editing and specificity values for ADAR1 and ADAR2, so a designer can ask for an isoform-selective or maximally clean guide.
• Structure learned from sequence: Guides are encoded as one-hot primary sequence, not folded structure, capturing non-canonical pairing that free-energy minimization misses.
• Activation-maximization module: The paired CNN ensemble optimizes guides for one well-screened target, and solved multi-site co-editing that exhaustive screening could not.

#Technical Details

The generalizable model is trained on the PolyTarget library, a biochemical high-throughput screen of 50,253 gRNAs across 5,643 distinct targets; earlier target-specific screens covered LRRK2 G2019S (~112,000 gRNAs) and the APP/BACE1 cleavage site (58,000 gRNAs). Splits are stratified by target (60/20/20 for diffusion, 80/20 for the supervised CNNs), so evaluation means unseen targets. Ensembles of 20 CNNs take concatenated one-hot target (41 nt) and gRNA (32–49 nt) sequences and reach Spearman's r of 0.58–0.67 on held-out targets. The diffusion model adds Gaussian noise to sequence encodings in [-1, 1] and reverses it with a denoising U-Net, trained 100 epochs with Adam at a 1e-4 learning rate. On a validation library of 37 target adenosines — 15 from PolyTarget and 22 unseen — diffusion beat rational heuristic designs on 35 of 37 targets for ADAR1 and 36 of 37 for ADAR2, and exceeded 50% on-target-only editing (a single edit, no bystanders on the same read) for 22 targets with ADAR1 and 30 with ADAR2, against 11 and 22 for ActMax. Neither code nor weights are released, and no model card, data card, or public copy of the screening libraries accompanies the preprint.

#Applications

The immediate use is therapeutic guide design: the paper's demonstration is a 294-design panel for the MECP2 R168X nonsense mutation behind Rett syndrome, conditioned for high on-target and low bystander editing, then scored with an XGBoost model for cross-species and allele-specific behavior. One design edited both human and mouse mutant transcripts above 51% while staying under 15% on wild-type and under 5% at every bystander adenosine. ActMax covers a different job: squeezing a heavily screened single target, including co-editing three adenosines in APP to install the K670G and M671V substitutions, which none of 58,000 screened designs achieved.

#Impact

DeepREAD moved ADAR guide design from a heuristic craft to a conditioned generative problem, showing that a diffusion model borrowed from image synthesis transfers to short structured RNA. Its influence is clearest in Shape's own pipeline: the company later paired it with Helix, a structure-aware predictive transformer, in a noisy-student distillation loop called DeepHelix, and the MECP2 R168X guide line advanced into preclinical work reporting roughly 70% brain-wide editing and extended survival in a mouse model. The work remains an unreviewed preprint and is proprietary — the authors are employees and inventors on related patents, and no artifacts are public, so outside groups cannot reproduce or extend it. Training also came from a biochemical assay with purified ADAR rather than cells, leaving competing RNA-binding proteins and endogenous ADAR levels out of what the model learned.