# CombFold weights

CombFold itself has no trainable weights. Pair and group structures are predicted
with pretrained AlphaFold-Multimer through ColabFold.

For a self-contained offline package, place the five official Multimer v3 files in:

```text
weight/alphafold/params/params_model_1_multimer_v3.npz
weight/alphafold/params/params_model_2_multimer_v3.npz
weight/alphafold/params/params_model_3_multimer_v3.npz
weight/alphafold/params/params_model_4_multimer_v3.npz
weight/alphafold/params/params_model_5_multimer_v3.npz
weight/alphafold/params/download_complexes_multimer_v3_finished.txt
```

Pass the directory above `params` to ColabFold:

```bash
colabfold_batch input.fasta output \
  --data weight/alphafold \
  --model-type alphafold2_multimer_v3
```

The SCNet adaptation reused weights stored outside this repository at:

```text
/public/home/scnb9biwet/wuxg/dl_binder_design-main/af2_initial_guess/model_weights/params
```