rna_dataset.py

import os
import random
import math
import torch
import json
from tqdm import tqdm
import numpy as np
from torch.utils.data import Dataset, DataLoader

import torch.nn.functional as F

from torch_geometric.data import Data
import pickle
from itertools import product
from nucleotides.load_nucleotide import get_standard_nucleotide
from scipy.spatial.transform import Rotation
import copy
from torch_geometric.data import Batch

N1 = 0
N9 = 0
N_dis_list = []

def continuous_cropping(data, rna_max_len):
    rna_len = data.x.size(0)
    start_idx = random.randint(0, rna_len - rna_max_len)
    end_idx = start_idx + rna_max_len
    data.x = data.x[start_idx:end_idx]
    if hasattr(data, 'pos'):
        data.pos = data.pos[start_idx:end_idx]
        # add struc_emb [0403 by TIANRUI]
        data.struc_emb = data.struc_emb[start_idx:end_idx]
        # add protenix_emb [1406 by YIMING]
        if hasattr(data, 'protenix_emb'):
            data.protenix_emb = data.protenix_emb[start_idx:end_idx]
    if hasattr(data, 'bead_pos'):
        data.bead_pos = data.bead_pos[start_idx:end_idx]
    edge_mask = (data.edge_index >= start_idx) & (data.edge_index < end_idx)
    edge_mask = edge_mask.all(dim=0)
    data.edge_index = data.edge_index[:, edge_mask] - start_idx
    data.edge_attr = data.edge_attr[edge_mask]
    data.pos_mask = data.pos_mask[start_idx:end_idx]
    data.res_id = data.res_id[start_idx:end_idx]
    data.residue_list = data.residue_list[start_idx:end_idx]
    return data

def spatial_cropping(data, rna_max_len):
    rna_len = data.x.size(0)
    if hasattr(data, 'pos') and data.pos.shape[1] == 9 and data.pos.dim() == 2:
        c_pos = data.pos[:, 3:6] # shape = [num_res, 3]
    elif hasattr(data, 'pos') and data.pos.dim() == 3 and data.pos.shape[1] == 3:
        c_pos = data.pos[:, 0]
    elif hasattr(data, 'pos') and data.pos.dim() == 3:
        c_pos = data.pos[:, 6] # c4' is the 7th in pos
    elif hasattr(data, 'bead_pos'):
        c_pos = data.bead_pos[:, 3:6] # shape = [num_res, 3]
    rand_center = c_pos[random.randint(0, rna_len - 1)].view(1, 3)
    dist = torch.linalg.vector_norm(c_pos - rand_center, dim=1) # shape = [num_res]
    topk = torch.topk(dist, rna_max_len, largest=False).indices
    topk = topk.sort().values

    ## update the node features
    data.x = data.x[topk]
    if hasattr(data, 'pos'):
        data.pos = data.pos[topk]
        # add struc_emb [0403 by TIANRUI]
        data.struc_emb = data.struc_emb[topk]
        # add protenix_emb [1406 by YIMING]
        if hasattr(data, 'protenix_emb'):
            data.protenix_emb = data.protenix_emb[topk]
    if hasattr(data, 'bead_pos'):
        data.bead_pos = data.bead_pos[topk]
    data.pos_mask = data.pos_mask[topk]
    data.res_id = data.res_id[topk]
    data.residue_list = [data.residue_list[i] for i in topk.tolist()]

    ## create a mapping from the original edge index to the new edge index
    edge_mapping = torch.ones(rna_len, dtype=torch.long) * -1
    edge_mapping[topk] = torch.arange(rna_max_len, dtype=torch.long)

    ## update the edge index
    new_edge_index = edge_mapping[data.edge_index.long()]
    edge_mask = (new_edge_index >= 0).all(dim=0) # shape = [num_edges]
    data.edge_index = new_edge_index[:, edge_mask]
    data.edge_attr = data.edge_attr[edge_mask]
    return data

def cropping(data, rna_max_len, spatial_crop_ratio=0.5):
    if random.random() < spatial_crop_ratio:
        data = spatial_cropping(data, rna_max_len)
    else:
        data = continuous_cropping(data, rna_max_len)
    return data

def random_rotate_pos(data, seed=None):
    """
    对 data.pos 进行统一随机旋转增强，返回新 Data
    - data.pos: [N, R, 3]，N 个节点，每个节点 R 个点
    """
    if seed is not None:
        rng = np.random.default_rng(seed)
    else:
        rng = np.random.default_rng()

    # 生成一个随机旋转矩阵
    rot = Rotation.random(random_state=rng)
    
    # 提取 pos 并旋转
    pos = data.pos.view(-1, 3)  # [N*R, 3]
    rotated_pos = rot.apply(pos)             # [N*R, 3]
    rotated_pos = torch.from_numpy(rotated_pos).to(data.pos.dtype)

    data.pos = rotated_pos.view_as(data.pos)  # reshape back to [N, R, 3]
    return data

def random_translation_pos(data, translation_scale=0.1):
    dtype = data.pos.dtype
    trans_aug = translation_scale * torch.randn(3, dtype=dtype)
    data.pos = data.pos + trans_aug.view(1, 1, 3)
    return data

def aug(data):
    # random_num = random.random()
    # if random_num < 1 / 3:
    data = random_rotate_pos(data)
    # elif random_num > 2 / 3:
    data = random_translation_pos(data)
    return data

class RNADataset(Dataset):
    def __init__(self, processed_path, rna_max_len=50, spatial_crop_ratio=0.5, disable_ss=False, mode='train', new_aa=True, new_res=False, fix_N_bug=False, struc_emb_path = '/home/hui007/rna/DiffRNA/data/RNAData20250323/struc_emb.pt', use_protenix_emb=True, protenix_emb_path='/home/hui007/protenix_0612/Protenix/protenix_embedding'):
        super(RNADataset, self).__init__()
        self.new_aa = new_aa
        self.new_res = new_res
        self.rna_max_len = rna_max_len
        self.residue_types = {'A':0, 'G':1, 'U':2, 'C':3}
        self.fix_N_bug = fix_N_bug
        self.use_protenix_emb = use_protenix_emb
        self.protenix_emb_path = protenix_emb_path
        if not os.path.exists(processed_path):
            raise FileNotFoundError(f'processed_path = `{processed_path}` not found! please preprocess it first!')
        
        ## load data
        assert processed_path.endswith('.npz')
        np_data = np.load(processed_path)
        self.cluster_mapping = pickle.loads(np_data['clstr'])
        # print(len(self.cluster_mapping), len(self.cluster_mapping[0]), len(self.cluster_mapping[0][0]), self.cluster_mapping[0][0])
        self.data_list = pickle.loads(np_data['data_list'])
        self.ss_list = {}
        for ss_id in np_data.keys():
            if ss_id.startswith('ss'):
                new_id = int(ss_id.split('_')[1])
                self.ss_list[new_id] = np_data[ss_id]

        self.rna_class_idx = {'[unclassed]': 0, 'solo': 1, 'protein-RNA': 2, 'DNA-RNA': 3}
        self.rnaclass = dict()
        
        # add struc_emb [0403 by TIANRUI]
        self.struc_emb = torch.load(struc_emb_path)

        self.mode = mode
        self.disable_ss = disable_ss
        self.pos_std = 20.3689
        
        self.spatial_crop_ratio = spatial_crop_ratio
        self.aa_mapping = get_standard_nucleotide(True, False)

        self.total_conf_num = len(self.data_list)
        assert self.total_conf_num == sum([len(idx_list) for cluster in self.cluster_mapping for idx_list in cluster])

        if self.mode == 'train':
            self.num_seq = sum([len(cluster) for cluster in self.cluster_mapping])
            self.seq_list = [inner for outer in self.cluster_mapping for inner in outer]

    def __len__(self):
        if self.mode == 'train':
            return self.num_seq
        elif self.mode in {'valid', 'test'}:
            return self.total_conf_num
        else:
            raise ValueError(f'Unsupported mode: {self.mode}')

    def len(self):
        return self.__len__()

    def get_idx_data(self, idx: int):
        data = self.data_list[idx]
        full_id = data['full_id']
        
        ## construct fully connected graph based on secondary structure
        num_res = len(data['data'])
        ss = self.ss_list[data['ss_id']]
        ss = torch.from_numpy(ss).to(torch.float)
        assert ss.shape[0] == num_res
        indices = np.mgrid[0:num_res, 0:num_res].reshape(2, -1)
        full_edge_index = torch.from_numpy(indices) # [2, N_edge]
        full_edge_attr = ss.reshape(-1, ss.size(-1)) # [N_edge, dim]

        rna_data = data['data']
        x = []
        residue_list = []
        pos = []
        pos_mask = []
        # add struc_emb [0403 by TIANRUI]
        struc_emb_list = []
        for _, res, atom_coords in rna_data:
            x.append(self.residue_types[res])
            residue_list.append(res)
            atom_list = atom_coords['atom_list']
            coord_list = atom_coords['coord_list']

            if self.new_aa:
                atom_mapping = self.aa_mapping[res][-1]
                aa_pos = torch.zeros((24, 3), dtype=torch.float)
                aa_mask = torch.ones((24,), dtype=torch.bool)
                for i, atom in enumerate(atom_list):
                    idx = atom_mapping.get(atom, None)
                    if idx is None:
                        continue
                    idx = idx[0]
                    aa_pos[idx] = torch.from_numpy(coord_list[i])
                    aa_mask[idx] = False
                pos.append(aa_pos)
                pos_mask.append(aa_mask)
                # add struc_emb [0403 by TIANRUI]
                struc_emb = self.struc_emb[res]
                struc_emb_list.append(struc_emb)

            else:
                if self.new_res:
                    atom_mapping = {'P': 0, 'C4\'': 1, 'N1': 2}
                    bead_pos = torch.zeros((3, 3), dtype=torch.float)
                    bead_pos_mask = torch.ones((3,), dtype=torch.bool)
                    for i, atom in enumerate(atom_list):
                        idx = atom_mapping.get(atom, None)
                        if idx is None:
                            continue
                        bead_pos[idx] = torch.from_numpy(coord_list[i])
                        bead_pos_mask[idx] = False
                    pos.append(bead_pos)
                    pos_mask.append(bead_pos_mask)
                else:
                    bead_idx = np.asarray([-1, -1, -1])
                    for i, atom in enumerate(atom_list):
                        if atom == 'P':
                            bead_idx[0] = i
                        elif atom == 'C4\'':
                            bead_idx[1] = i
                        else: 
                            if self.fix_N_bug:
                                if atom == 'N9' and (res == 'A' or res == 'G'):
                                    bead_idx[2] = i
                                elif atom == 'N1' and (res == 'C' or res == 'U'):
                                    bead_idx[2] = i
                            else:
                                if atom == 'N1' or atom == 'N9':
                                    bead_idx[2] = i
                    if (bead_idx == -1).any():
                        pos.append(torch.zeros((9,), dtype=torch.float))
                        pos_mask.append(True)
                    else:
                        bead_pos = coord_list[bead_idx].reshape(-1) # shape = [9]
                        pos.append(torch.from_numpy(bead_pos))
                        pos_mask.append(False)

        x = torch.LongTensor(x)
        x = F.one_hot(x, num_classes=4).to(torch.float)
        rna_class = self.rnaclass.get(full_id.split('-')[0], '[unclassed]')
        rna_class = self.rna_class_idx[rna_class]
        
        if self.new_aa:
            pos = torch.stack(pos, dim=0) # shape = [num_res, res_len, 3]
            pos_mask = torch.stack(pos_mask, dim=0) # shape = [num_res, res_len]
            data = Data(x=x, pos=pos, pos_mask=pos_mask, rna_seq_id=full_id, edge_index=full_edge_index, edge_attr=full_edge_attr, res_id=torch.arange(num_res), residue_list=residue_list)
            data['class'] = rna_class
            # add struc_emb [0403 by TIANRUI]
            struc_emb = torch.stack(struc_emb_list, dim=0)
            data['struc_emb'] = struc_emb
            # add protenix embedding [1206 by YIMING]
            if self.use_protenix_emb:
                protenix_emb = os.path.join(self.protenix_emb_path, f'{full_id}.pt')
                if os.path.exists(protenix_emb):
                    data['protenix_emb'] = torch.load(protenix_emb, map_location='cpu')
                else:
                    data['protenix_emb'] = torch.zeros(x.shape[0], 384, device='cpu')
        else:
            if self.new_res:
                pos_mask = torch.stack(pos_mask, dim=0)
            else:
                pos_mask = torch.BoolTensor(pos_mask)
            pos = torch.stack(pos, dim=0)
            data = Data(x=x, pos=pos, pos_mask=pos_mask, rna_seq_id=full_id, edge_index=full_edge_index, edge_attr=full_edge_attr, res_id=torch.arange(num_res), residue_list=residue_list)
            data['class'] = rna_class
        return data

    def __getitem__(self, idx):
        if self.mode == 'train':
            assert 0 <= idx < self.num_seq
            data_id = random.choices(self.seq_list[idx])[0]
            data = self.get_idx_data(data_id)
        elif self.mode in {'valid', 'test'}:
            data = self.get_idx_data(idx)
        else:
            raise ValueError(f'Unsupported mode: {self.mode}')
        
        if self.mode == 'test':
            if self.new_aa:
                ## center the data
                pos_unmask = ~data.pos_mask # shape = [num_res, res_len]
                centroid = data.pos.reshape(-1, 3).sum(dim=0, keepdim=True) / torch.clamp(pos_unmask.sum(), min=1) # shape = [1, 3]
                data.pos = (data.pos - centroid.unsqueeze(0)) * pos_unmask.unsqueeze(-1)
            else:
                if self.new_res:
                    ## center the data
                    pos_unmask = ~data.pos_mask # shape = [num_res, 3]
                    centroid = data.pos.reshape(-1, 3).sum(dim=0, keepdim=True) / torch.clamp(pos_unmask.sum(), min=1) # shape = [1, 3]
                    data.pos = (data.pos - centroid.view(1, 1, 3)) * pos_unmask.view(-1, 3, 1)
                else:
                    ## center the data
                    pos_unmask = ~data.pos_mask # shape = [num_res]
                    centroid = data.pos.reshape(-1, 3).sum(dim=0, keepdim=True) / torch.clamp((pos_unmask.sum() * 3), min=1) # shape = [1, 3]
                    data.pos = (data.pos - centroid.repeat(1, 3)) * pos_unmask.view(-1, 1)
            return data
        
        data_copy = copy.deepcopy(data)
        rna_len = data.x.size(0)
        if rna_len > self.rna_max_len:
            data = cropping(data, self.rna_max_len, self.spatial_crop_ratio)
            data_copy = cropping(data_copy, self.rna_max_len, self.spatial_crop_ratio)
        
        if self.new_aa:
            ## center the data
            pos_unmask = ~data.pos_mask # shape = [num_res, res_len]
            centroid = data.pos.reshape(-1, 3).sum(dim=0, keepdim=True) / torch.clamp(pos_unmask.sum(), min=1) # shape = [1, 3]
            data.pos = (data.pos - centroid.unsqueeze(0)) * pos_unmask.unsqueeze(-1) / self.pos_std

            pop_unmask_copy = ~data_copy.pos_mask
            centroid_copy = data_copy.pos.reshape(-1, 3).sum(dim=0, keepdim=True) / torch.clamp(pop_unmask_copy.sum(), min=1)
            data_copy.pos = (data_copy.pos - centroid_copy.unsqueeze(0)) * pop_unmask_copy.unsqueeze(-1) / self.pos_std
        
        else:
            if self.new_res:
                ## center the data
                pos_unmask = ~data.pos_mask # shape = [num_res, 3]
                centroid = data.pos.reshape(-1, 3).sum(dim=0, keepdim=True) / torch.clamp(pos_unmask.sum(), min=1) # shape = [1, 3]
                data.pos = (data.pos - centroid.view(1, 1, 3)) * pos_unmask.view(-1, 3, 1) / self.pos_std
            else:
                ## center the data
                pos_unmask = ~data.pos_mask # shape = [num_res]
                centroid = data.pos.reshape(-1, 3).sum(dim=0, keepdim=True) / torch.clamp((pos_unmask.sum() * 3), min=1) # shape = [1, 3]
                data.pos = (data.pos - centroid.repeat(1, 3)) * pos_unmask.view(-1, 1) / self.pos_std
        
        assert not torch.isnan(data.pos).any()
        
        data = aug(data)
        data_copy = aug(data_copy)

        return data, data_copy

class RNACollater:
    def __call__(self, batch):
        # batch: List of tuples (g1, g2)
        g1_list, g2_list = zip(*batch)

        g1_batch = Batch.from_data_list(g1_list)
        g2_batch = Batch.from_data_list(g2_list)

        # batch['max_seqlen'] = int((batch['ptr'][1:] - batch['ptr'][:-1]).max())
        g1_batch['edge_attr'] = g1_batch['edge_attr'].to(torch.float)
        g1_batch['edge_index'] = g1_batch['edge_index'].to(torch.long)
        g1_batch['class'] = torch.LongTensor(g1_batch['class'])
        # if hasattr(batch, 'res_id'):
        g1_batch['seq_pos'] = g1_batch.pop('res_id').to(torch.long)
        
        if hasattr(g1_batch, 'bead_pos'):
            g1_batch['gt_pos'] = g1_batch.pop('bead_pos')
        elif hasattr(g1_batch, 'pos'):
            g1_batch['gt_pos'] = g1_batch.pop('pos')

        # if self.new_res:
        #     batch['gt_pos'] = batch['gt_pos'].view(-1, 9)
        #     batch['pos_mask'] = batch.pos_mask.any(dim=-1)

        pos_unmask = ~g1_batch['pos_mask'].unsqueeze(-1)
        g1_batch['gt_pos'] = g1_batch['gt_pos'] * pos_unmask

        g2_batch['edge_attr'] = g2_batch['edge_attr'].to(torch.float)
        g2_batch['edge_index'] = g2_batch['edge_index'].to(torch.long)
        g2_batch['class'] = torch.LongTensor(g2_batch['class'])
        g2_batch['seq_pos'] = g2_batch.pop('res_id').to(torch.long)
        if hasattr(g2_batch, 'bead_pos'):
            g2_batch['gt_pos'] = g2_batch.pop('bead_pos')
        elif hasattr(g2_batch, 'pos'):
            g2_batch['gt_pos'] = g2_batch.pop('pos')
        pos_unmask = ~g2_batch['pos_mask'].unsqueeze(-1)
        g2_batch['gt_pos'] = g2_batch['gt_pos'] * pos_unmask

        return g1_batch, g2_batch

class RNACollater_v2:
    def add_noise(self, batch):
        noise = torch.randn(batch.gt_pos.shape) # shape = [\sum N_i, 9]
        batch['noise'] = noise
        
        pos_t = batch.gt_pos + 0.1 * noise # shape = [\sum N_i, 9]
        
        batch['pos'] = pos_t
        return batch

    def __call__(self, batch):
        # batch: List of tuples (g1, g2)
        g1_list, g2_list = zip(*batch)

        g1_batch = Batch.from_data_list(g1_list)
        g2_batch = Batch.from_data_list(g2_list)

        # batch['max_seqlen'] = int((batch['ptr'][1:] - batch['ptr'][:-1]).max())
        g1_batch['edge_attr'] = g1_batch['edge_attr'].to(torch.float)
        g1_batch['edge_index'] = g1_batch['edge_index'].to(torch.long)
        g1_batch['class'] = torch.LongTensor(g1_batch['class'])
        # if hasattr(batch, 'res_id'):
        g1_batch['seq_pos'] = g1_batch.pop('res_id').to(torch.long)
        
        if hasattr(g1_batch, 'bead_pos'):
            g1_batch['gt_pos'] = g1_batch.pop('bead_pos')
        elif hasattr(g1_batch, 'pos'):
            g1_batch['gt_pos'] = g1_batch.pop('pos')

        g1_batch = self.add_noise(g1_batch)

        # if self.new_res:
        #     batch['gt_pos'] = batch['gt_pos'].view(-1, 9)
        #     batch['pos_mask'] = batch.pos_mask.any(dim=-1)

        pos_unmask = ~g1_batch['pos_mask'].unsqueeze(-1)
        g1_batch['gt_pos'] = g1_batch['gt_pos'] * pos_unmask
        g1_batch['pos'] = g1_batch['pos'] * pos_unmask
        g1_batch['noise'] = g1_batch['noise'] * pos_unmask

        g2_batch['edge_attr'] = g2_batch['edge_attr'].to(torch.float)
        g2_batch['edge_index'] = g2_batch['edge_index'].to(torch.long)
        g2_batch['class'] = torch.LongTensor(g2_batch['class'])
        g2_batch['seq_pos'] = g2_batch.pop('res_id').to(torch.long)
        if hasattr(g2_batch, 'bead_pos'):
            g2_batch['gt_pos'] = g2_batch.pop('bead_pos')
        elif hasattr(g2_batch, 'pos'):
            g2_batch['gt_pos'] = g2_batch.pop('pos')

        g2_batch = self.add_noise(g2_batch)

        pos_unmask = ~g2_batch['pos_mask'].unsqueeze(-1)
        g2_batch['gt_pos'] = g2_batch['gt_pos'] * pos_unmask
        g2_batch['pos'] = g2_batch['pos'] * pos_unmask
        g2_batch['noise'] = g2_batch['noise'] * pos_unmask

        return g1_batch, g2_batch
    
if __name__ == '__main__':
    dataset = RNADataset('/home/hui007/rna/rna_repr/zhiyuan/train_data_final.npz')
    for i in range(10):
        print(dataset[i])
    loader = DataLoader(
            dataset,
            batch_size=4,
            shuffle=False,
            num_workers=0,
            drop_last=True,
            collate_fn=RNACollater()
        )
    for g1, g2 in loader:
        print(g1)
        print(g2)
        break