pair_alignment/dloaders/FullLenDset.py

#!/usr/bin/env python3
# -*- coding: utf-8 -*-
"""
Created on Thu Nov 30 12:39:48 2023


About:
======
Custom pytorch dataset object for giving pfam data to length-based models


outputs:
========
1. sample_unaligned_seqs: (1, L_seq, 2)

2. sample_aligned_mat: (1, L_align, d)
   > for pairHMM models: d = 3
     >> dim2 = 0: gapped ancestor
     >> dim2 = 1: gapped descendant
     >> dim2 = 2: categorically-encoded alignment (<pad>, M, I, D, <bos>, <eos>)
  
    > for feedforward head: d = 4
    >> dim2 = 0: descendant, under alignment-augmented alphabet (ins + A)
    >> dim2 = 1: categorically-encoded alignment (<pad>, M, I, D, <bos>, <eos>)
    >> dim2 = 2: m-indices, precalculated from alignment
    >> dim2 = 3: n-indices, precalculated from alignment
  
   > for neural pairHMM models: d = 5
     >> dim2 = 0: gapped ancestor
     >> dim2 = 1: gapped descendant
     >> dim2 = 2: categorically-encoded alignment (<pad>, M, I, D, <bos>, <eos>)
     >> dim2 = 3: m-indices, precalculated from alignment
     >> dim2 = 4: n-indices, precalculated from alignment
         
3. sample_time: 
    > (1,), if using one branch length per sample
    > returns None, otherwise
    
4. sample_idx: (1,)
    
use FullLenDset.retrieve_sample_names(sample_indices) to retrieve pairID,
  names of both sequences, and the pfam name


Data to be read:
=================
1. seqs_unaligned.npy: Numpy matrix of unaligned inputs; (B, L_align.max(), 2),
   where dim2 corresponds to-
    - (dim2=0): ungapped ancestor sequence
    - (dim2=0): ungapped descendant sequence

2. aligned_mats.npy: Numpy matrix of aligned inputs: (num_pairs, L_seq.max(), 4),
   where dim2 corresponds to-
    - (dim2=0): aligned ancestor sequence
    - (dim2=1): aligned descendant sequence
    - (dim2=2): m indexes (indices for ancestor alignment)
    - (dim2=3): n indexes (indices for descendant alignment)

3. metadata.tsv: [PANDAS DATAFRAME]
   > note: alignment length in this dataframe does NOT include 
     sentinel tokens!

4 pair-times.tsv: (B,)
  > plain .tsv file with two columns; no header and no index
  > first column is pairID
  > second column is time

"""
# general python
import numpy as np
import pandas as pd

# jax stuff
import jax 
from jax import numpy as jnp
from jax.tree_util import tree_map

# pytorch stuff
import torch
from torch.utils.data import Dataset, DataLoader,default_collate


###############################################################################
### pytorch collator   ########################################################
###############################################################################
def _default_collate_to_jax_array(mat):
    """
    kind of cumbersome, but conversion path is 
        tuple -> pytorch tensor -> numpy array -> jax array
    """
    pytorch_tensor = default_collate(mat)
    numpy_mat = pytorch_tensor.numpy()
    return jnp.array( numpy_mat )
    

def jax_collator(batch):
    """
    collator that can handle if time per sample is None
    
    B = number of samples in the batch
    L_seq = max length of the unaligned sequences (ancestor and descendant) + 2
    L_seq = max length of the aligned pairs + 2
    S = number of transitions; 4 here: M, I, D, START/END
    
    
    Returns
    -------
    collated_unaligned_seqs : ArrayLike, (B, L_seq, 2)
        > dim2[0]: ancestor sequence, encoded without gaps
        > dim2[1]: descendant sequence, encoded without gaps
    
    collated_aligned_mat : ArrayLike, (B, L_align, d)
        > for pairHMM models: d = 3
            >> dim2[0]: gapped ancestor
            >> dim2[1]: gapped descendant
            >> dim2[2]: categorically-encoded alignment (<pad>, M, I, D, <bos>, <eos>)
      
        > for feedforward head: d = 4
            >> dim2[0]: descendant, under alignment-augmented alphabet (ins + A)
            >> dim2[1]: categorically-encoded alignment (<pad>, M, I, D, <bos>, <eos>)
            >> dim2[2]: m-indices, precalculated from alignment
            >> dim2[3]: n-indices, precalculated from alignment
      
        > for neural pairHMM models: d = 5
            >> dim2[0]: gapped ancestor
            >> dim2[1]: gapped descendant
            >> dim2[2]: categorically-encoded alignment (<pad>, M, I, D, <bos>, <eos>)
            >> dim2[3]: m-indices, precalculated from alignment
            >> dim2[4]: n-indices, precalculated from alignment
    
    collated_times : ArrayLike, (B,) OR None
    
    collated_idx : ArrayLike, (B,)
    
    """
    # unpack batch
    out = zip(*batch)
    sample_unaligned_seqs, sample_aligned_mat, sample_time, sample_idx = out
    del out
    
    # handle unaligned_seqs, aligned_mat, and idx with default_collate
    collated_unaligned_seqs = _default_collate_to_jax_array( sample_unaligned_seqs )
    collated_aligned_mat = _default_collate_to_jax_array( sample_aligned_mat )
    collated_idx = _default_collate_to_jax_array( sample_idx )
    
    # handle time, which could be none
    if (sample_time[0] is not None):
        collated_times = _default_collate_to_jax_array( sample_time )
    
    elif (sample_time[0] is None):
        collated_times = None
        
    return (collated_unaligned_seqs, 
            collated_aligned_mat,
            collated_times,
            collated_idx)


###############################################################################
### some helpers   ############################################################
###############################################################################
def _remove_excess_padding(seqs, 
                          padding_tok: int):
    """
    trim excess padding
    """
    global_max_len = np.where(seqs != padding_tok,
                              True,
                              False).sum(axis=1).max()
    clipped_seqs = seqs[:, :global_max_len, ...]
    return clipped_seqs, global_max_len


def _add_padding_dim_1(mat,
                      padding_length: int,
                      padding_tok: int):
    """
    add padding to dim1 of matrix (usually length)
    """
    final_dtype = mat.dtype
    new_shape = (mat.shape[0], padding_length, mat.shape[2])
    padding = np.ones( new_shape, dtype =  final_dtype) * padding_tok
    padded_mat = np.concatenate( [mat, padding], axis=1)
    return padded_mat


def _pad_to_length_divisible_by_chunk_len(aligned_mat,
                                         padding_tok: int,
                                         chunk_length: int = 512):
    """
    to make sure seqs is divisible by chunk_length, may need to 
      pad with extra tokens
    
    this is used when padding alignment_mats for use with loss functions
      that use jax.lax.scan
    """
    global_max_len = aligned_mat.shape[1]
    
    num_chunks = 1
    while (chunk_length * num_chunks) < global_max_len:
        num_chunks += 1
    
    # add 1 for <bos>
    padding_length = ((chunk_length * num_chunks) - global_max_len) + 1
    
    final_aligned_mat = _add_padding_dim_1(mat = aligned_mat,
                                          padding_length = padding_length,
                                          padding_tok = padding_tok)
    
    return final_aligned_mat, padding_length
    



###############################################################################
### functions to load raw data   ##############################################
###############################################################################
def _load_aligned_mats(data_dir, 
                      split, 
                      pred_model_type,
                      emission_alphabet_size,
                      toss_alignments_longer_than = None, 
                      gap_idx = 43,
                      bos_idx = 1,
                      eos_idx = 2):
    """
    alignment encoding:
        
        <pad> = 0
        M = 1
        I = 2
        D = 3
        <bos> = 4
        <eos> = 5
    """
    ### load data
    with open(f'{data_dir}/{split}_aligned_mats.npy','rb') as f:
        mat = np.load(f)
       
        
    ### if alignments are longer than toss_alignments_longer_than, 
    ###   then toss the samples
    if toss_alignments_longer_than:
        eos_locs = np.argwhere(mat[...,0] == eos_idx)
        idxes_to_keep = eos_locs[ eos_locs[:, 1] <= toss_alignments_longer_than ][:, 0]
        mat = mat[idxes_to_keep, :, :]
        del eos_locs
        
        if len(idxes_to_keep) == 0:
            raise RuntimeError(f"no samples to keep from {split}!")
            
    else:
        idxes_to_keep = None
    
    
    ### encode alignment state; bos and eos are shifted to the end, 
    ###   to match overleaf document
    # <pad> = 0
    # M = 1
    # I = 2
    # D = 3
    # <bos> = 4
    # <eos> = 5
    alignment = np.zeros(mat.shape[:2], dtype=np.int8)  # (B, L)
    gapped_seqs = mat[...,[0,1]] # (B, L, 2)
    
    # matches
    mask = ((gapped_seqs >= 3) & (gapped_seqs <= 22)).sum(axis=2) == 2
    alignment[mask] = 1
    del mask
    
    # ins: ancestor is gap
    alignment[gapped_seqs[..., 0] == gap_idx] = 2
    
    # del: descendant is gap
    alignment[gapped_seqs[..., 1] == gap_idx] = 3
    
    # bos, eos
    alignment[mat[..., 0] == bos_idx] = 4
    alignment[mat[..., 0] == eos_idx] = 5
    
    
    ### model-specific transformations, concatenation
    ### feedforward: add 20 to insert sites in descendant, toss ancestor
    if pred_model_type == 'feedforward':
        # zero-padded items
        gapped_anc = gapped_seqs[...,0] #(B, L)
        gapped_desc = gapped_seqs[...,1] #(B, L)
        
        # insert sites are where ancestor = gap char; add 20 here (in place)
        ins_pos = np.argwhere( gapped_anc == gap_idx ) #(B,)
        gapped_desc[ ins_pos[:,0], ins_pos[:,1] ] += emission_alphabet_size #(B, L, 3)
        zero_padded_mat = np.stack([gapped_desc, alignment], axis=-1) # (B, L, 2)
        del gapped_anc, gapped_desc, ins_pos, alignment
        
        # -9 padded items
        neg_nine_padded_mat = mat[...,[-2,-1]] # (B, L, 2)
    
    
    ### pairHMM: concatenate zero-padding matrix; toss negative nine-padding matrix
    elif pred_model_type in ['pairhmm_indp_sites',
                             'pairhmm_frag_and_site_classes',
                             'pairhmm_nested_tkf']:
        zero_padded_mat = np.concatenate([gapped_seqs, alignment[...,None]], axis=-1) # (B, L, 3)
        neg_nine_padded_mat = None
    
    
    ### neural pairHMM: concatenate both
    elif pred_model_type == 'neural_hmm':
        zero_padded_mat = np.concatenate([gapped_seqs, alignment[...,None]], axis=-1) # (B, L, 3)
        neg_nine_padded_mat = mat[...,[-2,-1]] # (B, L, 2)
        
    return zero_padded_mat, neg_nine_padded_mat, idxes_to_keep



def _load_unaligned(data_dir, 
                   split, 
                   idxes_to_keep=None):
    with open(f'{data_dir}/{split}_seqs_unaligned.npy','rb') as f:
        mat = np.load(f)
    
    if (idxes_to_keep is not None):
        mat = mat[idxes_to_keep, :, :]
    
    return mat


def _load_metadata(data_dir, 
                  split, 
                  idxes_to_keep=None):
    cols_to_keep = ['pairID',
                    'ancestor',
                    'descendant',
                    'pfam', 
                    'anc_seq_len', 
                    'desc_seq_len', 
                    'alignment_len',
                    'num_matches',
                    'num_ins',
                    'num_del']
    
    df = pd.read_csv( f'./{data_dir}/{split}_metadata.tsv', 
                     sep='\t', 
                     index_col=0,
                     usecols=cols_to_keep) 
    
    if (idxes_to_keep is not None):
        df = df.iloc[idxes_to_keep]
    
    return df


###############################################################################
### functions to postprocess   ################################################
###############################################################################
def _postprocess_aligned_mats(zero_padded_aligned_mats_lst,
                             neg_nine_padded_aligned_mats_lst,
                             divisible_by_chunk_length: bool,
                             chunk_length: int = 512,
                             seq_padding_idx: int = 0,
                             align_padding_idx: int = -9):
    """
    zero_padded_aligned_mats_lst: list of matrices to concatenate, which use 
        zero as the padding token
    
    neg_nine_padded_aligned_mats_lst: list of matrices to concatenate, which  
        use -9 as the padding token
    
    divisible_by_chunk_length [BOOL]: True if using scanned version of
        loss function; False otherwise
    
    chunk_length [INT=512]: used for lengths in scan and determining number of 
        jit-compiled functions; if not provided, use 512
    
    seq_padding_idx, align_padding_idx: what the padding tokens are
    """
    # concat
    zero_padded_aligned_mats = np.concatenate(zero_padded_aligned_mats_lst,
                                              axis=0)
    if neg_nine_padded_aligned_mats_lst is not None:
        neg_nine_padded_aligned_mats = np.concatenate(neg_nine_padded_aligned_mats_lst,
                                                      axis=0)
    else:
        neg_nine_padded_aligned_mats = None
    
    del zero_padded_aligned_mats_lst, neg_nine_padded_aligned_mats_lst
    
    
    ### first half; adjust gapped ancestor and descendant seqs
    # remove excess padding
    out = _remove_excess_padding(seqs = zero_padded_aligned_mats, 
                                padding_tok = 0)
    final_mat, align_max_len_without_padding = out
    del out
    
    # if you want this to be divisible by chunk_length, may need to 
    #   add more padding tokens (0)
    if divisible_by_chunk_length:
        out = _pad_to_length_divisible_by_chunk_len(aligned_mat = final_mat,
                                                   padding_tok = 0,
                                                   chunk_length = chunk_length)
        final_mat, extra_padding_to_add = out
        del out
        
        
    ### second half; adjust precomputed alignment indices
    if neg_nine_padded_aligned_mats is not None:
        # remove excess padding; already calculated the length for this, so
        #   just reuse that
        second_half = neg_nine_padded_aligned_mats[:, :align_max_len_without_padding, :]
        
        # if you want this to be divisible by chunk_length, may need to 
        #   add more padding tokens (-9); again, already calculated length for this
        if divisible_by_chunk_length:
            second_half = _add_padding_dim_1(mat = second_half,
                                            padding_length = extra_padding_to_add,
                                            padding_tok = -9)
        
        final_mat = np.concatenate([final_mat, second_half], axis=-1)
    
    return final_mat


def _postprocess_unaligned_seqs(in_lst,
                               seq_padding_idx: int = 0):
    unaligned_seqs = np.concatenate(in_lst, axis=0)
    unaligned_seqs, _ = _remove_excess_padding(seqs = unaligned_seqs, 
                                              padding_tok = seq_padding_idx)
    return unaligned_seqs
    
    
def _postprocess_metadata(in_lst):
    metadata_df = pd.concat(in_lst)
    metadata_df = metadata_df.reset_index(drop=True)
    
    return metadata_df




###############################################################################
### Main dataset object   #####################################################
###############################################################################
class FullLenDset(Dataset):
    def __init__(self, 
                 data_dir: str, 
                 split_prefixes: list, 
                 pred_model_type: str,
                 use_scan_fns: bool,
                 t_per_sample: bool,
                 toss_alignments_longer_than = None,
                 chunk_length: int = 512,
                 emission_alphabet_size: int = 20,
                 seq_padding_idx: int = 0,
                 align_padding_idx: int = -9,
                 gap_idx: int = 43):
        """
        Load pairwise alignments and metadata
        

        Arguments
        ----------
        data_dir : str
            Where data is located
            
        split_prefixes : List[str]
            prefixes of the datasets to include
        
        pred_model_type : ['pairhmm_indp_sites', 'pairhmm_frag_and_site_classes', 
                           'pairhmm_nested_tkf', 'feedforward', 'neural_hmm']
            what the broad classification of the model is; changes behaviors here
        
        use_scan_fns : bool
            If True, use jax.lax.scan implementation of likelihood functions
        
        emission_alphabet_size : int
            4 if DNA, 20 if proteins
        
        t_per_sample : bool
            True if you want to read a branch length per sample, False otherwise
            
        chunk_length : int, optional
            Pad samples in increments of this
            DEFAULT VALUE: 512
        
        toss_alignments_longer_than : int, None
            Max alignment length to keep, if desired
            DEFAULT VALUE: None
          
        seq_padding_idx : int, optional
            DEFAULT VALUE: 0
        
        align_padding_idx : int, optional
            DEFAULT VALUE: -9
        
        gap_idx : int, optional
            DEFAULT VALUE: 43

        
        Attributes created
        -------------------
          self.unaligned_seqs
          self.aligned_mat
          self.names_df
          self.times
          self.emit_counts
          self.global_seq_max_length
          self.global_align_max_length
            > global_align_max_length is divisible by chunk_length 
              if using scan version of functions
        """
        ###############
        ### read data #
        ###############
        # always read
        zero_padded_aligned_mats_lst = []
        neg_nine_padded_aligned_mats_lst = []
        unaligned_seqs_lst = []
        metadata_lst = []
        self.emit_counts = np.zeros( (emission_alphabet_size,) )
        
        if emission_alphabet_size == 20:
            counts_suffix = 'AAcounts'
        
        elif emission_alphabet_size == 4:
            counts_suffix = 'NuclCounts'
            
        # optionally read
        if t_per_sample:
            times_lst = []
        
        for split in split_prefixes:
            ### aligned inputs: alignment, and precalculated (m,n) indices
            ###   remove any samples with alignments greater than toss_alignments_longer_than
            out = _load_aligned_mats(data_dir = data_dir, 
                                    split = split, 
                                    toss_alignments_longer_than = toss_alignments_longer_than, 
                                    pred_model_type = pred_model_type,
                                    gap_idx = gap_idx,
                                    emission_alphabet_size = emission_alphabet_size)
            zero_padded_mat, neg_nine_padded_mat, idxes_to_keep = out
            del out
            
            zero_padded_aligned_mats_lst.append( zero_padded_mat )
            if neg_nine_padded_mat is not None:
                neg_nine_padded_aligned_mats_lst.append( neg_nine_padded_mat )
                
            del zero_padded_mat, neg_nine_padded_mat
            
            
            ### unaligned inputs (the sequences themselves)
            ###   remove any samples with alignments greater than toss_alignments_longer_than
            unaligned_seqs = _load_unaligned(data_dir = data_dir, 
                                            split = split,
                                            idxes_to_keep = idxes_to_keep)
            
            unaligned_seqs_lst.append(unaligned_seqs)
            del unaligned_seqs
            
            
            ### metadata
            meta_df = _load_metadata(data_dir = data_dir, 
                                    split = split,
                                    idxes_to_keep = idxes_to_keep)
            metadata_lst.append(meta_df)
            del meta_df
            
            
            ### counts of amino acids
            with open(f'{data_dir}/{split}_{counts_suffix}.npy','rb') as f:
                self.emit_counts += np.load(f)
            
            
            ### (optional) time; assume time is in same order as samples in
            ###   metadata
            if t_per_sample:
                times = pd.read_csv(f'{data_dir}/{split}_pair-times.tsv', 
                                    sep='\t',
                                    header=None,
                                    names=['pairID','time'],
                                    index_col=None)

                if (idxes_to_keep is not None):
                    times = times.iloc[idxes_to_keep]

                times_lst += times['time'].tolist()
                del times
                
                
        #################
        ### postprocess #
        #################
        # matrix of alignment info
        lst2 = None if len(neg_nine_padded_aligned_mats_lst) == 0 else neg_nine_padded_aligned_mats_lst
        self.aligned_mat = _postprocess_aligned_mats(zero_padded_aligned_mats_lst = zero_padded_aligned_mats_lst,
                                                    neg_nine_padded_aligned_mats_lst = lst2,
                                                    divisible_by_chunk_length = use_scan_fns,
                                                    chunk_length = chunk_length,
                                                    seq_padding_idx = seq_padding_idx,
                                                    align_padding_idx = align_padding_idx)
        self.global_align_max_length = self.aligned_mat.shape[1]
        del zero_padded_aligned_mats_lst, neg_nine_padded_aligned_mats_lst, lst2
        
        # ungapped seqs
        self.unaligned_seqs = _postprocess_unaligned_seqs(in_lst = unaligned_seqs_lst,
                                                         seq_padding_idx = seq_padding_idx)
        self.global_seq_max_length = self.unaligned_seqs.shape[1]
        del unaligned_seqs_lst
        
        # metadata
        self.names_df = _postprocess_metadata(in_lst = metadata_lst)
        del metadata_lst
        
        # (optional) time
        if t_per_sample:
            self.times = np.array(times_lst) #(B,)
            del times_lst
        
        else:
            self.times = None
        
    def __len__(self):
        return self.aligned_mat.shape[0]

    def __getitem__(self, idx):
        sample_unaligned_seqs = self.unaligned_seqs[idx, ...]
        sample_aligned_mat = self.aligned_mat[idx, ...]
        
        if self.times is not None:
            sample_time = self.times[idx]
        else:
            sample_time = None
        
        sample_idx = idx
        
        return (sample_unaligned_seqs, 
                sample_aligned_mat, 
                sample_time, 
                sample_idx)
    
    def retrieve_sample_names(self, idxes):
        # used the list of sample indices to query the original names_df
        return self.names_df.iloc[idxes]