GraFold: code, datasets, results and checkpoints for “A Graph-Attention Fusion Network with Thermodynamic Priors for RNA Secondary Structure Prediction”
Description
Everything needed to check or rebuild the results of the GraFold manuscript: the pipeline that builds each dataset from public sources, trains both stages, runs the baselines and produces the tables and figures; the second implementation used for the replication on an 8 GB GPU and for the later experiments; the pre-registration documents written before those experiments; the stored results, audits and logs; the trained checkpoints; and the derived datasets. Baseline code and weights from other groups are not redistributed; the scripts fetch them from their authors. Start with README.md.
Files
Steps to reproduce
This is a computational study; no instruments or reagents were used. All steps are carried out by the released scripts. 1. Sources: bpRNA-1m with the TR0/VL0/TS0 splits of Singh et al. (2019); ArchiveII with the splits of Szikszai et al. (2022); Rfam 12.2 and 14.10 seed alignments; RCSB PDB (search of 13 September 2026). The scripts download them. 2. Derived datasets. PDB: entries split at deposition date 1 January 2022; resolution at most 3.5 Å (NMR kept); one chain per RNA entity; 20-500 nt; CD-HIT-EST at 80 % identity; test chains at least 80 % identical to a fine-tuning chain removed; base pairs annotated from coordinates with RNApolis (cis Watson-Crick, |i-j| >= 4, at least four pairs). Result: 570 fine-tuning and 259 test structures. Rfam cross-family set: 10,791 seed sequences from 1,036 families present in Rfam 14.10 but not in 12.2. Augmentation: 15,000 sequences made by mutating 10-30 % of the residues of TR0 sequences, labelled with CONTRAfold 2.02. Counts at every filtering step are in the audit logs. 3. Independence audits: training and fine-tuning data were compared with every evaluation set by exact match, BLASTN identity (80 %), CD-HIT-EST-2D (80 %), Rfam family (Infernal) and, for PDB structures, US-align (TM-score 0.5). Flagged items are listed, and results are also given without them. 4. Model and training: pairing probabilities and stacking energies from ViennaRNA 2.7.2 (Turner 2004); model in PyTorch 2.11.0. Stage 1: 80 epochs on TR0 plus the augmented sequences (AdamW, learning rate 3e-4, focal + BCE loss); checkpoint and threshold chosen on VL0. Stage 2: 20 epochs on the PDB fine-tuning set. Decoding: maximum-weight matching (NetworkX). 5. Baselines: RNAfold, MXfold2, SPOT-RNA, UFold, RNAformer, BPfold, RiNALMo and GCNfold were run from their authors' code and weights on the same sequences; RNA-FM with a structure head trained here. Their code and weights are not included; the scripts fetch them. 6. Scoring: strict base-pair precision, recall and F1 per sequence, macro-averaged; 95 % bootstrap confidence intervals; paired Wilcoxon signed-rank tests with Bonferroni or Holm correction. 7. Hardware: main pipeline on one NVIDIA A100 (Google Colab), about 165 GPU-hours; replication and later experiments on one NVIDIA RTX 5050 laptop GPU (8 GB); added baselines on NVIDIA L4 and T4. 8. To reproduce: create the conda environments (environment.yml, environment-bio.yml). Main results: run 1_pipeline_colab/GraFold_Colab_A100.ipynb with the "balanced" profile. Replication and later experiments: the run chains in 2_pipeline_local/scripts; each needs its file from 3_preregistration, where the decision rules were written before the runs. README.md gives the commands.
Institutions
- Manav Rachna UniversityHaryana, Faridabad