GraFold

Description

RNA Secondary Structure Prediction

Steps to reproduce

Step-by-Step Running Guide Track A — fast first result (~1–2 h) !python grafold_colab.py --mode setup !python grafold_colab.py --mode download !python grafold_colab.py --mode preprocess --max_len 384 --max_train 5000 --rfam_versions "" !python grafold_colab.py --mode train --nb 8 --gd 128 --max_len 384 --accum 1 --epochs 40 --no_amp !python grafold_colab.py --mode evaluate --nb 8 --gd 128 --max_len 384 WARNING Watch the first val F1 lines. If they stay 0.0000 or print ABORT, switch to the stable width model --nb 4 --gd 192 (for train, finetune AND evaluate). 10.2 Track B — full paper-faithful run !python grafold_colab.py --mode download !python grafold_colab.py --mode preprocess --max_len 384 --rfam_versions 12.3,14.10 !python grafold_colab.py --mode augment --augment 15000 --max_len 384 !python grafold_colab.py --mode train --nb 8 --gd 128 --max_len 384 --accum 1 --epochs 80 --no_amp !python grafold_colab.py --mode evaluate --nb 8 --gd 128 --max_len 384 !python grafold_colab.py --mode pdb --pdb_test 120 --pdb_ft 782 --pdb_cutoff 2021-01-01 --max_len 384 !python grafold_colab.py --mode finetune --nb 8 --gd 128 --max_len 384 --accum 1 !python grafold_colab.py --mode evaluate --nb 8 --gd 128 --max_len 384 Four rules that prevent the common failures 4. Keep --nb and --gd identical across train / finetune / evaluate. 5. Always pass --accum 1 (default 32 undertrains here). 6. Always pass --max_len 384 (omitting → 512 → OOM). 7. Use --no_amp for stability (FP16 can diverge to NaN). . Analysis Experiments Ablation study Trains five cumulative variants (baseline → +outer → +energy → +GAT → +cross) and reports F1 on bpRNA-TS0. !python grafold_colab.py --mode ablation --nb 4 --gd 192 --max_len 320 --accum 1 --epochs 10 --max_train 4000 Leave-one-family-out Trains on all-but-one ArchiveII family, tests the held-out family; reports per-family F1. !python grafold_colab.py --mode lofo --nb 4 --gd 192 --max_len 320 --accum 1 --epochs 10 Pseudoknot prediction F1 on pseudoknot (crossing) base pairs only, for structures containing them. No training. !python grafold_colab.py --mode pseudoknot --pk_set test_ts0 --nb 8 --gd 128 --max_len 384 Computational efficiency Inference time + peak memory vs length (100→1000 nt). Paper: ~340 ms for 300 nt on an A100. !python grafold_colab.py --mode efficiency --nb 8 --gd 128 NOTE Ablation trains 5 models and LOFO ~10 — keep --epochs modest. Efficiency and pseudoknot need no training.

Institutions

• Manav Rachna University

  Haryana, Faridabad

Categories

Licence