Loading…
Designs RNA and DNA aptamers against protein targets by backpropagating binding and anti-binding objectives through a frozen all-atom predictor.
No providers recorded yet. Browse all providers
EGFR and HER2 are close relatives in the ErbB receptor family, and their extracellular domain III surfaces look much alike. A design method that only maximizes predicted confidence in the intended complex has no way to express "and not that one" — off-target discrimination can only be applied afterwards, by discarding candidates. NACraft makes the off-target part of the objective instead: one candidate sequence is scored in two molecular contexts at once, a binding loss pulling it toward EGFR while a sign-flipped anti-binding loss pushes it away from HER2, both gradients landing on the sequence every step.
NACraft is a training-free framework for designing RNA and DNA aptamers, released as a preprint in August 2026 by Valhalla Technology with New York University, Peking University, Yonsei University and The Chinese University of Hong Kong. It represents an aptamer of length L as a relaxed logit tensor over the four RNA and four DNA nucleotides, masks the alphabet not in use, and backpropagates structural losses computed on the distogram head of Boltz-1 into those logits. Boltz-1's weights never move; the optimization moves the sample. NACraft introduces no trained parameters of its own, and pointing it at a new target means writing a design-context YAML rather than collecting data and fitting a model.
This inverts a structure predictor the way BindCraft and BoltzDesign1 do, but for a molecule those methods optimize only as context. Crossing over needed a geometric vocabulary that fits nucleotides: contact masks are defined on C1' reference atoms rather than the protein Cα/Cβ conventions.
nucleic_acid_class field masks the logit tensor to
the four RNA or four DNA nucleotides, so both polymer types run through identical code.A straight-through estimator connects discrete nucleotide identities to the differentiable search, and the pseudo-sequence handed to Boltz-1 is a scheduled blend of logits, relaxed probabilities and hard assignments — 30 warm-up, 100 exploration, 100 annealing and 10 low-temperature iterations, following BoltzDesign1's staged strategy. Optimized sequences are then materialized by Boltz-1 diffusion and rediversified with NA-MPNN.
De novo design produced 6,000 RNA candidates across five therapeutic targets — B7-H3, PD-L1, CD3δ, TNFR1 and FGFR2 — at 20 to 50 nucleotides, with best AF3 ipTM of 0.63 to 0.88 per target and the fraction above ipTM 0.60 ranging from 1.08% for CD3δ to 41.50% for FGFR2. On NA-12, six protein–RNA and six protein–DNA complexes deposited in the PDB after 13 January 2023, 7,200 scored candidates reached target-wise maximum ipTM of 0.93 for RNA and 0.91 for DNA, NA-MPNN redesign raising the maximum for all twelve. In the EGFR-over-HER2 campaign, 69.44% of 1,800 candidates scored higher ipTM for EGFR, RNA discriminating more sharply than DNA (82.56% versus 56.33%). Under matched AF3 rescoring NACraft exceeded the diffusion-based ODesign on target-level median ipTM in 10 of 11 NA-12 targets and 17 of 20 protein target–length settings.
Aptamers serve as capture reagents, biosensor elements, diagnostic probes and therapeutic candidates, and are normally discovered by SELEX — enrichment from randomized libraries, which gives little control over which surface gets bound. NACraft offers a structure-directed complement: de novo mode proposes binders for a target with no known aptamer, similarity-guided mode remodels the interface of an existing SELEX-derived sequence while a soft prior keeps it recognizable, and target-selective mode addresses a homologous receptor that must be spared. Running it end to end is demanding — a Linux host with an NVIDIA GPU, separately downloaded Boltz-1 assets and the NA-MPNN design checkpoint, both public downloads, and a full AlphaFold 3 installation whose parameters are the one genuine gate: they must be requested from Google DeepMind and used under its terms.
NACraft extends structure-model hallucination, developed almost entirely on protein binders, to nucleic acids, and locates the interesting part in the objective interface rather than the backbone: once binding, anti-binding and sequence priors are separable terms over multiple evaluation contexts, distinct aptamer-engineering problems become weight settings. The evidence is entirely computational. Every reported number is an AF3 confidence metric — ipTM, pLDDT, iPAE — and the authors state plainly that these do not establish affinity, specificity or activity, so no candidate should be read as a confirmed binder; a single final verifier also risks propagating its own biases. The implementation covers canonical RNA and DNA nucleotides only, with no modified bases or explicit ligand modality. The code is MIT-licensed and the Zenodo deposit CC BY 4.0, while the preprint is CC BY-NC and awaits peer review.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.