QuickFold

Roughly 10,000 RNA structures sit in the Protein Data Bank against millions of known sequences, a scarcity that puts RNA folding in a different regime from the protein problem AlphaFold 2 solved, and that shapes the RNA predictors that followed: RhoFold+ and NuFold compensate with larger trunks and a homology search that dominates their wall-clock cost. QuickFold takes the opposite position, pushing nearly all of its prior knowledge into two frozen single-sequence signals and keeping the part that must be fitted on scarce structural data deliberately small.

Those signals are a per-nucleotide embedding from the frozen RNA-FM language model — the same class of feature RhoFold+ uses — and a ViennaRNA base-pair probability matrix, injected as a pair feature through a zero-initialised linear map so that the roughly 92% of entries with negligible pairing probability contribute exactly zero and the bias fires only on confident base pairs. Everything downstream is an 8.9M-parameter trained trunk, and no alignment is computed at any point.

QuickFold was developed at Sanofi Digital R&D as the terminal state of an experiment in how models get built: over 297 iterations a general-purpose LLM agent proposed, implemented, trained and evaluated one change at a time under a fixed per-iteration budget on a single GPU, with a human supplying visual structural judgement and scope decisions rather than the architecture. The authors frame the paper as a case study in feedback-driven development rather than as a new state of the art; QuickFold's claim is parity within seed noise with the strongest open-source RNA predictors at a fraction of their inference cost.

#Key Features

• MSA-free by construction: the deployed variant takes only the query sequence, after a late change in the development run removed the alignment input at no cost in accuracy.
• Two frozen conditioning signals: a language-model embedding supplies evolutionary context and a thermodynamic base-pair matrix supplies secondary structure, so the trained network never has to rediscover either from 18,675 training chains.
• Chemically valid atom placement: heavy atoms are never regressed directly. A predicted rigid frame is applied to an idealised per-nucleotide template — the medoid conformation over training residues, which avoids the bond-length shrinkage averaging would introduce — with a small glycosidic-χ head rotating the base ring, so all learned error lives in the inter-residue frames.
• Seconds-scale folding: a full prediction takes about 3.5 seconds per target, of which the folding trunk itself is 0.09 seconds and the rest is an optional van der Waals declash step.
• Calibrated confidence: trained distogram, pLDDT and PAE heads accompany every prediction, and pLDDT selects among independently trained seeds.

#Technical Details

The forward pass runs sequence → featurizer → ten trunk blocks → a four-iteration structure module → heads and atom placement, with one recycle, at a single-representation width of 256 and a pair width of 128. Each trunk block applies pair-biased self-attention with a transition, an OuterProductMean, and the gated triangle-multiplicative pair update imported from AlphaFold 2; the structure module is an invariant point attention stack maintaining a per-residue rigid frame. The ten-term objective keeps frame-aligned point error, distogram, pLDDT, PAE and a glycosidic-χ term active throughout while ramping soft-TM, backbone continuity, bond-angle and clash terms in only over the final 40% of the budget — geometric and global-fold penalties destroy an unformed fold and help only once a rough one exists. Training data is a redundancy-reduced set of single-chain PDB RNAs of 16–500 nucleotides split by deposition date, with the 80 test chains deposited from 2024 onward so the split post-dates the baselines' training cutoffs too. The run was from scratch on one NVIDIA L40S under a 480-minute wall-clock budget, with EMA-averaged weights deployed.

On the 80-target held-out set QuickFold reaches lDDT 0.600 ± 0.016 and TM-score 0.306 across three seeds, against 0.591 / 0.283 for NuFold and 0.531 / 0.272 for RhoFold+ — differences inside seed noise. Inference averages 3.5 seconds per target versus 20.0 for NuFold and 274.9 for RhoFold+, excluding MSA search. The trained trunk is 8.9M parameters against 27.2M for RhoFold+ and 94.8M for NuFold; counting the frozen RNA-FM encoder, QuickFold carries 108.4M at inference. Global fold remains the weak axis at TM ≈ 0.31, and the authors attribute the lDDT plateau near 0.60 to a data ceiling rather than an architectural one.

#Applications

QuickFold suits work where many RNA targets must be folded cheaply and no useful homology exists: sweeping structures across a transcriptome, triaging orphan or synthetic constructs before committing to experimental structure determination, or generating starting conformations for simulation. Its speed and small trunk also make it a practical control in methods work — a single-sequence baseline retrainable end to end within hours on one commodity GPU.

#Impact

QuickFold's contribution is methodological. It shows that a compact, carefully conditioned network can reach the accuracy band of established RNA predictors, and that the cost profile that follows — seconds rather than minutes, with no alignment step — changes which experiments are affordable. Its provenance is the more unusual claim: the architecture was assembled by an agent working against a measured objective rather than designed by hand, and the authors are candid that this route reaches the level of existing methods comfortably but did not push past them. No code, weights or checkpoints have been released; the paper notes that the output of an autonomous loop must be cleaned and inspected before release, so the model is a reported result rather than a runnable tool. The work is a preprint and has not been peer reviewed.