protenix-rna

Protenix-RNA

Protenix-RNA is a Protenix fine-tuned PyTorch checkpoint optimized for RNA structure prediction. The current checkpoint was selected by the EMA validation lDDT-complex best metric at training step 16,999 and is distributed as a native Protenix checkpoint for the Protenix codebase, not as a transformers.AutoModel package.

Files

The checkpoint is a torch.load(..., weights_only=False) dictionary with keys model, optimizer, scheduler, and step. The stored step is 16999.

Training Summary

• Base model: protenix_base_default_v1.0.0
• Fine-tuning data: local RNA fine-tune split from outputs/rna_finetune_full
• Validation split size: 478 PDB IDs
• Training crop size: 384 tokens
• Validation max tokens: 768
• RNA MSA: enabled
• Protein MSA and templates: disabled
• EMA decay: 0.999
• Optimizer for the current continuation: Aurora
• Selection metric: rna_finetune_val/ema0.999_lddt/complex/best.avg, maximize
• Current selection metric value: 0.772730 at step 16,999
• Full eval settings: seed 42, bf16, N_sample=5, N_step=20, N_cycle=4, max_n_token=768
• Full eval size after token filtering: 1,490 target rows from 195 PDB IDs

Full RNA Evaluation

Higher is better for lDDT, TM-score, and pLDDT. Lower is better for loss.

The full comparison table below was produced for the previous step-12,999 checkpoint. The current step-16,999 checkpoint has updated validation metrics in checkpoint_info.json; the long full comparison has not been rerun yet.

These values come from a full comparison run against protenix_base_default_v1.0.0 using the same RNA validation setup and saved predictions for the step-12,999 checkpoint.

Top-100 Structure Comparison

The top-100 set is selected from the full Protenix-RNA eval by tm_score_c1prime_best, then matched against base-model predictions from the same validation set. This subset is useful for inspecting best-case RNA behavior; it is not an unbiased dataset average.

The following PyMOL-rendered collage shows rank-1 predicted structures from representative top-100 targets, colored by atom pLDDT stored in the mmCIF B-factor field.

Checkpoint Selection Trace

This checkpoint was selected from the EMA validation loop by lDDT-complex best at step 16,999.

Usage

Download the checkpoint and point Protenix at it with --load_params_only true:

hf download LiteFold/protenix-rna \
  checkpoints/best_ema_0.999.pt \
  --local-dir ./protenix-rna

Example evaluation invocation inside the Protenix checkout:

LOAD_CHECKPOINT_PATH=./protenix-rna/checkpoints/best_ema_0.999.pt \
VAL_MAX_N_TOKEN=768 \
VAL_LIMIT=-1 \
N_SAMPLE=5 \
N_STEP=20 \
N_CYCLE=4 \
./run_rna_latest_full_eval_tm_dump.sh

For direct loading:

import torch

ckpt = torch.load("checkpoints/best_ema_0.999.pt", map_location="cpu", weights_only=False)
state_dict = ckpt["model"]
step = ckpt["step"]

Limitations

This is a research checkpoint specialized for the RNA fine-tuning setup above. It has not been converted into a standalone Transformers model and should be evaluated with the same Protenix code/configuration family used for training.