AlphaFold3 data_transforms 模块 make_atom14_masks 函数的作用是为蛋白质的每个氨基酸残基构建一个更密集的原子表示（14个维度而不是37个维度），并生成相应的掩码和映射。这是蛋白质结构预测任务中的一个重要步骤，因为它允许模型专注于更关键的原子，同时忽略一些不太重要的原子，从而提高计算效率和预测精度。
源代码：
def make_atom14_masks(protein):"""Construct denser atom positions (14 dimensions instead of 37)."""restype_atom14_to_atom37 = []restype_atom37_to_atom14 = []restype_atom14_mask = []for rt in rc.restypes:atom_names = rc.restype_name_to_atom14_names[rc.restype_1to3[rt]]restype_atom14_to_atom37.append([(rc.atom_order[name] if name else 0) for name in atom_names])atom_name_to_idx14 = {name: i for i, name in enumerate(atom_names)}restype_atom37_to_atom14.append([(atom_name_to_idx14[name] if name in atom_name_to_idx14 else 0)for name in rc.atom_types])restype_atom14_mask.append([(1.0 if name else 0.0) for name in atom_names])# Add dummy mapping for restype 'UNK'restype_atom14_to_atom37.append([0] * 14)restype_atom37_to_atom14.append([0] * 37)restype_atom14_mask.append([0.0] * 14)restype_atom14_to_atom37 = torch.tensor(restype_atom14_to_atom37,dtype=torch.int32,device=protein["aatype"].device,)restype_atom37_to_atom14 = torch.tensor(restype_atom37_to_atom14,dtype=torch.int32,device=protein["aatype"].device,)restype_atom14_mask = torch.tensor(restype_atom14_mask,dtype=torch.float32,device=protein["aatype"].device,)protein_aatype = protein['aatype'].to(torch.long)# create the mapping for (residx, atom14) --> atom37, i.e. an array# with shape (num_res, 14) containing the atom37 indices for this proteinresidx_atom14_to_atom37 = restype_atom14_to_atom37[protein_aatype]residx_atom14_mask = restype_atom14_mask[protein_aatype]protein["atom14_atom_exists"] = residx_atom14_maskprotein["residx_atom14_to_atom37"] = residx_atom14_to_atom37.long()# create the gather indices for mapping 
