AF3 make_msa_feat函数解读

AlphaFold3  data_transforms 模块的 make_msa_feat 函数的作用是为多序列比对（MSA）数据创建和拼接特征，生成 msa_feat 和 target_feat 特征，这些特征将被用于后续的模型输入。

源代码：

@curry1
def make_msa_feat(protein):
    """Create and concatenate MSA features."""
    # Whether there is a domain break. Always zero for chains, but keeping for
    # compatibility with domain datasets.
    has_break = torch.clip(
        protein["between_segment_residues"].to(torch.float32), 0, 1
    )
    aatype_1hot = make_one_hot(protein["aatype"], 21)

    target_feat = [
        torch.unsqueeze(has_break, dim=-1),
        aatype_1hot,  # Everyone gets the original sequence.
    ]

    msa_1hot = make_one_hot(protein["msa"], 23)
    has_deletion = torch.clip(protein["deletion_matrix"], 0.0, 1.0)
    deletion_value = torch.atan(protein["deletion_matrix"] / 3.0) * (
            2.0 / np.pi
    )

    msa_feat = [
        msa_1hot,
        torch.unsqueeze(has_deletion, dim=-1),
        torch.unsqueeze(deletion_value, dim=-1),
    ]

    if "cluster_profile" in protein:
        deletion_mean_value = torch.atan(
            protein["cluster_deletion_mean"] / 3.0
        ) * (2.0 / np.pi)
        msa_feat.extend(
            [
                protein["cluster_profile"],
                torch.unsqueeze(deletion_mean_value, dim=-1),
            ]
        )

    if "extra_deletion_matrix" in protein:
        protein["extra_has_deletion"] = torch.clip(
            protein["extra_deletion_matrix"], 0.0, 1.0
        )
        protein["extra_deletion_value"] = torch.atan(
            protein["extra_