RNA folding kinetics model that predicts the full distribution of first passage times from a few simulated examples in a single forward pass.
No providers recorded yet. Browse all providers
How quickly an RNA reaches its functional fold is a separate question from what that fold looks like. Kinetic simulators such as Kinfold answer it by running Markov-chain Monte Carlo trajectories across the secondary-structure landscape and recording first passage times — the time each trajectory needs to reach a target structure for the first time. Thousands of trajectories are needed before the cumulative distribution function (CDF) of those times becomes statistically meaningful, and runtime grows exponentially with chain length. That cost keeps kinetic analysis off the table for long RNAs and for design loops that must score many candidates.
KinPFN, from the AutoML group at the University of Freiburg with collaborators at the University of Vienna, attacks the cost rather than the simulator. It is a prior-data fitted network: a transformer trained once on synthetic data to perform amortized Bayesian inference. Given a small set of real first passage times for a new RNA as context, it returns the posterior predictive distribution over folding times — and therefore the whole CDF — in one forward pass, with no gradient updates and no per-molecule fitting. The work was released as a bioRxiv preprint in October 2024 and published at ICLR 2025.
What makes the model general is what it deliberately never sees. KinPFN takes no sequence, no structure, no length, and no energy parameters, only the observed times. That leaves it agnostic to whatever process generated the numbers, so one checkpoint slots behind any kinetic simulator and the same machinery transfers to unrelated biological distributions. It sits alongside RNAformer, the same group's MSA-free secondary structure predictor, and shares its inference paradigm with PerturbPFN and GPFN.
KinPFN is a 4.86 million parameter transformer with 8 layers, an embedding size of 256, a hidden size of 512, and 4 attention heads, emitting a discretized Riemann distribution over 1,000 buckets. Training used 4.3 million synthetic datasets (1,000 epochs of 86 steps at batch size 50), each holding M = 1,400 first passage times sampled from multi-modal Gaussian families and log-encoded over the range 10⁻⁶ to 10¹⁵. Each dataset is split at a random cutoff so the model learns to condition on any amount of context, minimizing the prior-data negative log-likelihood. Hyperparameters were tuned with NePS, and the final model trained in roughly five hours on a single A40 GPU.
Evaluation used a new test set of 635 random RNA sequences of 15 to 147 nucleotides with 1,000 Kinfold simulations each. From 25 context times upward, KinPFN records the lowest negative log-likelihood of every method compared — 1.2435 against 1.2559 for kernel density estimation and 1.3612 for a two-component Gaussian mixture — while consuming 2.5% of the simulator compute; at only 10 context times, kernel density estimation remains ahead on mean absolute error and the Kolmogorov-Smirnov statistic. For a 76-nucleotide yeast tRNA-Phe, 50 context times reproduced the ground-truth CDF, cutting Kinfold CPU time from roughly 2,686 minutes to 170.
The immediate beneficiaries are groups that already run kinetic folding simulations: riboswitch and aptamer designers comparing how efficiently sequence variants reach a shared minimum-free-energy structure, synthetic biologists tuning co-transcriptional folding, and RNA therapeutics teams treating folding efficiency as a design constraint. Because the speedup is multiplicative across candidates, KinPFN makes kinetics-aware design loops tractable where exponential simulator runtimes ruled them out. The feature-free interface also invites reuse outside kinetics — any assay yielding a sample of times or counts whose full distribution is expensive to measure.
KinPFN is the first deep learning approach to RNA folding kinetics, a corner of RNA modeling that structure prediction had largely passed by, and it shows that the prior-data fitted network recipe developed for tabular data and hyperparameter optimization carries into molecular biophysics when real training data is unobtainable. Its accuracy is bounded by how much simulation a user will run: approximations improve steadily with more context, so the method trades compute for fidelity rather than eliminating the simulator. Code, data, and the pretrained checkpoint are released under Apache-2.0.
Much of this page is generated or calculated automatically. Flag anything that looks off and we will re-run it.