RiboGRAM Archives
RiboGRAM Archives This release contains the source code, model checkpoints, benchmark datasets, and example outputs for the study RiboGRAM learns multi-scale non-coding RNA grammar for structural and regulatory inference. RiboGRAM is a self-supervised RNA representation model for learning multi-scale and position-resolved sequence organization in non-coding RNAs (ncRNAs). The release supports sequence embeddings, attention visualization, ncRNA category and family classification, RNA-protein interaction (RPI) prediction, RNA-RNA interaction (RRI) prediction, and RNA-chromatin interaction (RCI) prediction. It also includes the RiboGRAM-3D inference package for MSA-free RNA tertiary-structure modeling. Project archive and web resource: https://www.ultrarnalab.com/ribogram.php Release contents The release is organized into three ZIP archives with a combined size of approximately 53.4 GB. Archive Contents Size 1. Source Code.zip RiboGRAM and RiboGRAM-3D source packages 9.02 GB 2. Model Availability.zip Pretrained and task-specific model checkpoints 22.08 GB 3. Dataset Availability.zip Training, evaluation, structure, and interaction datasets 22.29 GB Source code RiboGRAM The RiboGRAM.zip package is a compact Python inference implementation. It includes: infer.py: command-line entry point and public inference functions; ribogram/: model, alphabet, convolution, attention, and checkpoint-loading modules; checkpoints/: the base RiboGRAM checkpoint and the two ncRNA classification heads; test.ipynb: an executed inference notebook; test_samples.json: provenance and labels for the example records; test_outputs/: example embeddings, attention maps, classifications, and interaction predictions. Supported tasks: Task Command Required checkpoint Sequence embeddings embedding --architecture RiboGRAM RiboGRAM.pkl PLM-CDS embeddings embedding --architecture PLM-CDS PLM-CDS.pkl PLM-mRNA embeddings embedding --architecture PLM-mRNA PLM-mRNA.pkl Attention maps attention --architecture RiboGRAM RiboGRAM.pkl 13-category ncRNA classification ncrna13 RiboGRAM.pkl, ribogram_ncrna13_head.pt 88-family ncRNA classification ncrna88 RiboGRAM.pkl, ribogram_ncrna88_head.pt RNA-protein interaction interaction --architecture RiboGRAM-RPI RiboGRAM-RPI.pkl Expanded RNA-protein interaction interaction --architecture RiboGRAM-RPIplus RiboGRAM-RPI_plus.pkl RNA-RNA interaction interaction --architecture RiboGRAM-RRI RiboGRAM-RRI.pkl RNA-chromatin interaction interaction --architecture RiboGRAM-RCI RiboGRAM-RCI.pkl The RiboGRAM/README.md file inside the source package contains the complete command-line reference and Python API examples. RiboGRAM-3D The RiboGRAM-3D.zip package contains the RNA tertiary-structure inference runtime, including: infer_structure.py and inference/infer.py; the packaged epoch-500 structure checkpoint; data/DRfold2/ with test CIF structures, test PT files, and train/test manifests; data loaders, model modules, the RNA structure module, and the local openfold-3/ source tree. RiboGRAM-3D requires one additional STAR RNA language-model backbone checkpoint as an input to the inference script. This backbone is a separate model used for sequence features; it is not a RiboGRAM checkpoint and is not included in this release. Obtain a compatible backbone separately and pass its path with --backbone-ckpt. The package also does not include the full DRfold2 training CIF/PT collection. Its training manifests are lookup tables for that external collection. See the RiboGRAM-3D_new/README.md, inference/README.md, and data/DRfold2/README.md files inside the source package for environment setup, checkpoint preparation, and structure-data details. Model checkpoints The Model Availability.zip package contains: Pretrained models RiboGRAM.pkl: ncRNA-pretrained RiboGRAM backbone; PLM-CDS.pkl: CDS-pretrained control model; PLM-mRNA.pkl: mRNA-pretrained control model. Classification heads ribogram_ncrna13_head.pt: 13-category ncRNA classification head; ribogram_ncrna88_head.pt: 88-family ncRNA classification head. Interaction models RiboGRAM-RPI.pkl: RNA-protein interaction model; RiboGRAM-RPI_plus.pkl: expanded RNA-protein interaction model; RiboGRAM-RRI.pkl: RNA-RNA interaction model; RiboGRAM-RCI.pkl: RNA-chromatin interaction model. The backbone and interaction checkpoints are large files, approximately 3.4 GB each. GPU inference is recommended. Checkpoint files use PyTorch/Python serialization and should be loaded only from a trusted release. Datasets The Dataset Availability.zip package contains the datasets used in the manuscript analyses. Nine-species RNA sequence collection Complete cDNA, CDS, and ncRNA collections are provided for: Homo sapiens (GRCh38); Mus musculus (GRCm39); Rattus norvegicus (mRatBN7.2); Danio rerio (GRCz11); Arabidopsis thaliana (TAIR10); Drosophila melanogaster (BDGP6.32); Caenorhabditis elegans (WBcel235); Saccharomyces cerevisiae (R64-1-1); Escherichia coli K-12 MG1655 (ASM584v2). This section also includes the stratified UMAP sample nine_species_umap_data_sample_5000.zip and the human ncRNA-versus-CDS and ncRNA-versus-mRNA classification datasets. Pretraining datasets RiboGRAM_PRETRAINING_ncRNA.zip: ncRNA pretraining corpus; ensembl_CDS.zip: CDS control corpus; ensembl_mRNA.zip: mRNA control corpus. The manuscript describes the RiboGRAM corpus as approximately 40 million non-redundant ncRNA sequences assembled from RNAcentral 25.0, Rfam 15.0, and Ensembl release 115. ncRNA representation-analysis datasets 13-Class _ncRNA_Classification_Dataset.csv; 88-Family_ncRNA_classification_dataset.csv; Rfam_v15_seed_sequences_full.csv; Rfam_v15_seed_sequences_top200_per_family.csv. RNA structure datasets ArchiveII.zip; bpRNA-1m_TR0_VL0.zip; bpRNA-1m_TR0_VL0_TS0.zip; TrainSetA-TestSetB.zip; RNA_tertiary_structure_prediction.zip. The RiboGRAM-3D benchmark contains 41 held-out RNA chains ranging from 27 to 363 nt. The manuscript describes 9,446 processed training chains derived from the DRfold2 RNA structure collection. RNA-target interaction datasets RPI-Core.zip and RPI-Scale.zip: RNA-protein interaction data; MirTarRAW.zip and RISE-derived.zip: RNA-RNA interaction data; RNA_DNA_dataset.zip: RNA-chromatin/RNA-DNA interaction data. Dataset construction, filtering, and train/validation/test splits are described in Supplementary Tables 14-16 of the manuscript. Quick start RiboGRAM inference Extract RiboGRAM.zip and run the commands from the directory containing infer.py. The supplied installer was validated with Python 3.8.18, PyTorch 1.13.1+cu116, and an NVIDIA A100 GPU. The installer is a Bash script; on Windows, use WSL or install the listed dependencies manually. Place the checkpoints required for the selected task in checkpoints/. The base RiboGRAM checkpoint and the two classification heads are already included in the source package. The other checkpoints are provided in Model Availability.zip. bash requirements.shpython infer.py embedding \ --architecture RiboGRAM \ --sequence rna-1=GGAGCUCAGCCUUCACUGCG \ --layer 12 \ --device auto \ --output outputs/RiboGRAM_embeddings.pt Generate the manuscript attention examples: python infer.py attention \ --architecture RiboGRAM \ --case all \ --layer 12 \ --head 20 \ --symmetrize \ --plot \ --output-dir outputs/attention For classification and interaction examples, see the command table above, RiboGRAM/README.md, and the executed test.ipynb. RiboGRAM-3D inference Extract RiboGRAM-3D.zip, obtain a compatible STAR RNA language-model backbone checkpoint separately, and run: python inference/infer.py \ --inputs /path/to/input.pt \ --backbone-ckpt /path/to/STAR-RNA-backbone.pkl \ --output-dir /path/to/output The structure checkpoint included in the package is used by the inference configuration. The run is resumable by default, and the output manifest records the commands and random seeds. See RiboGRAM-3D_new/inference/README.md for multi-sample inference and checkpoint preparation. Input conventions RNA is uppercased and T is converted to U; protein sequences are uppercased; DNA is uppercased and U is converted to T; whitespace inside sequences is removed; unsupported symbols are mapped to ; RPI input: RNA protein ; RRI input: RNA1 RNA2 ; RCI input: RNA DNA reverse-complement(DNA) . Reproducibility RiboGRAM uses a 12-layer hybrid encoder with a hidden dimension of 1,280 and 20 attention heads. The encoder integrates parallel convolutional branches with kernel sizes of 1, 3, 5, 7, and 9 nt and Rotary Position Embeddings (RoPE), and is pretrained using a masked language modeling (MLM) objective. To facilitate transparent reuse and independent validation, this archive provides the complete pretraining corpora for RiboGRAM and the CDS/mRNA control models, source code for the supported RiboGRAM and RiboGRAM-3D inference workflows, pretrained and task-specific model checkpoints, benchmark datasets used in the study, and representative test inputs with corresponding reference outputs. Together, these resources enable users to trace data provenance, run and evaluate the released models, and verify checkpoint loading and output formats without repeating the computationally intensive pretraining process. Project links RiboGRAM Archives: https://www.ultrarnalab.com/ribogram.php Manuscript: RiboGRAM learns multi-scale non-coding RNA grammar for structural and regulatory inference License and third-party data The supplied materials do not specify one project-level license. Software, model checkpoints, and datasets may be subject to different terms. Users must check the applicable license and