what is dist_mat_loss? ang_to_nm and rotation augmentation are not applied to motif-scaffolding training.

[YiyuHong]

Hi,

(1) I see there is a dist_mat_loss that always output nan when I running a motif scaffolding training with a small dataset.
Can the loss be just ignored? the default value 0.6nm was used for the training? I can't find this loss in the paper.

        # Add mask to only account for pairs that are closer than thr in ground truth
        max_dist = self.cfg_exp.loss.thres_aux_2d_loss
        if max_dist is None:
            max_dist = 1e10
        pair_mask_thr = gt_pair_dists < max_dist  # [*, n, n]
        total_pair_mask = pair_mask * pair_mask_thr  # [*, n, n]

        # Compute loss
        den = torch.sum(total_pair_mask, dim=(-1, -2)) - nres
        dist_mat_loss = torch.sum(
            (gt_pair_dists - pred_pair_dists) ** 2 * total_pair_mask, dim=(-1, -2)
        )  # [*]
        dist_mat_loss = dist_mat_loss / den  # [*]

(2) max_dist_boundary is set as 1.0nm default, Refering to the paper, it says 30A.
what exact value did you used to train the models?

Thanks in advance.

[YiyuHong]

There is a bug when training motif scaffolding, in the following line should add x_1 = ang_to_nm(x_1)

proteina/proteinfoundation/nn/motif_factory.py

Line 336 in 15e5b5e

x_1 = batch["coords"][:,:,1,:] # [b, n, 3]

Also, no random rotation augmentation is applied when training motif-scaffolding