Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in [1].
RNA-FrameFlow: Flow Matching for de novo 3D RNA Backbone Design
Authors: Rishabh Anand, Chaitanya K. Joshi, Alex Morehead, Arian Rokkum Jamasb, Charles Harris, Simon V Mathis, Kieran Didi, Rex Ying, Bryan Hooi, Pietro Lio
TMLR 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We introduce RNA-Frame Flow, the first generative model for 3D RNA backbone design. We build upon SE(3) flow matching for protein backbone generation and establish protocols for data preparation and evaluation to address unique challenges posed by RNA modeling. ... Additionally, we define a suite of evaluation metrics to measure whether the generated RNA structures are globally self-consistent (via inverse folding followed by forward folding) and locally recover RNA-specific structural descriptors. The most performant version of RNA-Frame Flow generates locally realistic RNA backbones of 40-150 nucleotides, over 40% of which pass our validity criteria as measured by a selfconsistency TM-score 0.45... Experiments: 3D RNA structure dataset. RNAsolo (Adamczyk et al., 2022) is a recent dataset... Evaluation metrics: We evaluate our models for unconditional RNA backbone generation... We generate 50 backbones for target lengths sampled between 40 and 150 at intervals of 10. We then compute the following indicators of quality for these backbones: Validity (sc TM 0.45)... Diversity... Novelty... Local structural measurements... Table 1: Unconditional RNA backbone generation. We evaluate the performance of RNAFrame Flow for multiple values for denoising steps NT. ... Table 2: Local structural metrics. Earth Mover s Distance for local structural measurements compared to ground truth measurements from RNAsolo. ... Table 3: Impact of data preparation strategies. ... Figure 4: Local structural metrics from 600 generated backbone samples, compared to random Gaussian point cloud as a sanity check. |
| Researcher Affiliation | Collaboration | 1Yale University 2University of Cambridge (UK) 3Lawrence Berkeley National Laboratory 4Prescient Design, Genentech, Roche 5University of Oxford (UK) 6National University of Singapore |
| Pseudocode | No | The paper describes the methodology, including steps for representing RNA backbones as frames, SE(3) Flow Matching, and imputation of non-frame atoms. For example, Section A.4 describes 'Imputing Non-frame Atoms from Torsion Angles' as a procedural description. However, none of these descriptions are formatted as structured pseudocode blocks or explicitly labeled algorithms within the main text or appendices. |
| Open Source Code | Yes | Open-source code: github.com/rish-16/rna-backbone-design |
| Open Datasets | Yes | RNAsolo (Adamczyk et al., 2022) is a recent dataset of RNA 3D structures extracted from isolated RNAs, protein-RNA complexes, and DNA-RNA hybrids from the Protein Data Bank (as of January 5, 2024). |
| Dataset Splits | No | RNAsolo (Adamczyk et al., 2022) is a recent dataset of RNA 3D structures extracted from isolated RNAs, protein-RNA complexes, and DNA-RNA hybrids from the Protein Data Bank (as of January 5, 2024). The dataset contains 14,366 structures at resolution 4 Å (1 Å = 0.1nm). We select sequences of lengths between 40 and 150 nucleotides (5,319 in total) as we envisioned this size range contains structured RNAs of interest for design tasks. ... We train generative models on all RNA structures of length 40-150 nucleotides from RNAsolo (Adamczyk et al., 2022). The paper mentions using a subset of the RNAsolo dataset for training but does not provide specific train/validation/test splits (e.g., percentages or counts) or reference predefined splits for reproducibility. |
| Hardware Specification | Yes | We train for 120K gradient update steps on four NVIDIA Ge Force RTX 3090 GPUs for 18 hours with a batch size B = 28. |
| Software Dependencies | No | The paper describes the architecture of the flow model and components like Invariant Point Attention layers, Transformer encoder layers, and an MLP head. It mentions using Adam W optimizer. However, it does not specify version numbers for any software libraries (e.g., PyTorch, TensorFlow, CUDA) or programming languages used for implementation. |
| Experiment Setup | Yes | We use 6 IPA blocks in our flow model, with an additional 3-layer torsion predictor MLP that takes in node embeddings from the IPA module. Our final model contains 16.8M trainable parameters. We use Adam W optimizer with learning rate 0.0001, β1 = 0.9, β2 = 0.999. We train for 120K gradient update steps on four NVIDIA Ge Force RTX 3090 GPUs for 18 hours with a batch size B = 28. Each batch contains samples of the same sequence length to avoid padding. Further hyperparameters are listed in Appendix A.1. Invariant Point Attention (IPA) Atom embedding dimension Dh 256 Hidden dimension Dz 128 Number of blocks 6 Query and key points 8 Number of heads 8 Key points 12 Transformer Number of heads 4 Number of layers 2 Torsion Prediction MLP Input dimension 256 Hidden dimension 128 Schedule Translations (training / sampling) linear / linear Rotations (training / sampling) linear / exponential Number of denoising steps NT 50 |