Multi-Modal Foundation · LiteFold
Protenix-RNA
A Protenix-lineage PyTorch model fine-tuned for RNA three-dimensional structure prediction.
What Protenix-RNA is
Protenix-RNA is a PyTorch biomolecular structure prediction checkpoint developed by LiteFold. Originating from the base model protenix_base_default_v1.0.0, it is specifically fine-tuned to predict the three-dimensional structures of ribonucleic acid (RNA) molecules.
Accurate computational modelling of RNA tertiary structures is essential for understanding gene expression, catalytic RNA functions, and RNA-targeting therapeutics. Protenix-RNA adapts biomolecular structure prediction capabilities directly towards RNA-specific modelling tasks, improving upon generic biomolecular predictions.
The checkpoint is distributed in native Protenix format, selected using an exponential moving average (EMA) validation metric assessing the local distance difference test (lDDT) across molecular complexes.
How it works
Protenix-RNA is fine-tuned from protenix_base_default_v1.0.0 using the Aurora optimizer and an exponential moving average decay of 0.999. The model utilizes RNA multiple sequence alignments (MSAs), whilst protein MSAs and structural templates are explicitly disabled during training.
Training is conducted with a crop size of 384 tokens and evaluated with a maximum length of 768 tokens using bf16 precision, with sampling parameters set to 5 samples, 20 steps, and 4 cycles. The model was fine-tuned on an RNA split derived from the LiteFold/PDB index and selected at step 16,999 based on maximizing the average best complex lDDT score.
Key facts
- Base model
- protenix_base_default_v1.0.0
- Framework
- PyTorch
- Training data
- LiteFold/PDB index (local RNA fine-tune split)
- Training crop size
- 384 tokens
- Maximum validation tokens
- 768 tokens
- Selected training step
- 16,999
- Optimizer
- Aurora
What you can use it for
- RNA Tertiary Structure Prediction
- Predicting three-dimensional coordinates and conformations of RNA chains from sequence and RNA multiple sequence alignments.
- Complex Structure Evaluation
- Generating and ranking structural candidates using local distance difference test (lDDT) and TM-score metrics for RNA-containing complexes.
- Per-Residue Confidence Estimation
- Extracting predicted local distance difference test (pLDDT) confidence metrics, stored in the mmCIF B-factor field, to assess regional model reliability.
Run it here
We list this model but cannot run it yet: no pipeline_tag — needs a human.
Everything on this page still applies. Running it needs a loader for this repository, not a bigger machine.
Strengths and limits
Good at
- Delivers substantial improvements over base Protenix in RNA metrics, reaching an EMA validation complex lDDT of 0.7727 at step 16,999.
- Demonstrates high structural concordance on top-performing targets, achieving a TM-score C1' rank-1 of 0.9602 on the top-100 evaluation set.
- Optimised specifically for RNA by incorporating RNA MSAs whilst eliminating unneeded protein MSAs and structural templates.
Watch out for
- Requires execution within the native Protenix repository codebase rather than being compatible with the standard Hugging Face transformers pipeline.
- Constrained to a training crop size of 384 tokens and a maximum evaluated token length of 768 tokens.
- Excludes protein MSAs and templates, restricting its usage profile to the specific RNA fine-tuning setup.
Inputs and outputs
You provide
Sequence or text
A protein or nucleotide sequence, or free text — whichever this model was trained on.
You get back
Mean-pooled representation, one vector per input.
Licence
Check the licence before commercial use · apache-2.0
The apache-2.0 licence carries terms worth reading before you depend on it.
Questions
Can I use Protenix-RNA commercially?
Check the licence before commercial use. The apache-2.0 licence carries terms worth reading before you depend on it.
Can Protenix-RNA be loaded using Hugging Face AutoModel?
No. The checkpoint is distributed as a native PyTorch dictionary containing model, optimizer, scheduler, and step keys, and it requires the native Protenix codebase to run.
What is the token length limit for this model?
The model is trained with a crop size of 384 tokens and configured for an evaluation maximum of 768 tokens.
What metrics were used to select the primary checkpoint?
The checkpoint was selected at training step 16,999 by maximizing the exponential moving average (EMA 0.999) best complex lDDT metric on the validation split.
Written from Protenix-RNA's own documentation (model card). Statements citing a figure that does not appear there are removed before publication. Licence and running details come from our registry.