Modelpublic

Xu-AI4Science/MARRI

sourceHugging Facecc-by-4.0updated 12d agoView on Hugging Face

0likes

Model Card

MARRI model checkpoints

Trained checkpoints for MARRI (mechanism-aware multimodal framework for RNA--RNA interaction prediction). All checkpoints share the same architecture (RNA-FM + RNet-2D dual-tower encoder, opening branch, hybridization-gain MLP, axial cross-attention classifier) and the same backbone configuration -- RNA-FM fine-tuned, RNet-2D frozen, epoch-wise dynamic negative resampling -- differing only in random seed and/or training data split.

Code: https://github.com/GainGod-Xu/MARRI Interpretability/attention export for the headline model: https://huggingface.co/datasets/Xu-AI4Science/MARRI-interpretability

Each .pth file is a raw state_dict() for the interaction model (not a full training checkpoint with optimizer state), loadable via model.load_state_dict(torch.load(path, map_location=device)).

Main-matrix checkpoints (N-clean dataset, random 8:1:1 split)

File	Seed	Test ROC-AUC	Best epoch	Notes
`dyn_ft-rnafm_seed42/best_rna_physics_model.pth`	42	0.7530	23	Headline model, reported as the main MARRI result throughout the paper
`dyn_ft-rnafm_seed43/best_rna_physics_model.pth`	43	0.7409	15	Seed-repeat run, same recipe
`dyn_ft-rnafm_seed44/best_rna_physics_model.pth`	44	0.7342	18	Seed-repeat run, same recipe

Mean +/- SD over these three seeds: ROC-AUC 0.7427 +/- 0.0095 (Table 1 in the paper).

Leakage-resistant held-out generalization checkpoints (seed 42, own data split each)

Same recipe, each retrained from scratch on its own leakage-resistant held-out split (a separate positive/negative pool per split, not the main N-clean dataset -- see Methods/Supplementary for split construction).

File	Held-out split	Test ROC-AUC	Best epoch
`generalization_interaction_pair_seed42/best_rna_physics_model.pth`	Unseen RNA--RNA pair	0.7422	14
`generalization_transcript_seed42/best_rna_physics_model.pth`	Unseen RNA/gene (most stringent)	0.6630	13
`generalization_interaction_locus_seed42/best_rna_physics_model.pth`	Unseen local interaction locus	0.7502	19

These three correspond to Table 4 ("Generalization under leakage-resistant held-out splits") in the paper.