Spaces:
Running
on
T4
Running
on
T4
# utility functions for dealing with contigs during hallucination | |
import numpy as np | |
import random, copy, torch, geometry, os, sys | |
from kinematics import xyz_to_t2d | |
def parse_range_string(el): | |
''' Splits string with integer or integer range into start and end ints. ''' | |
if '-' in el: | |
s,e = el.split('-') | |
s,e = int(s), int(e) | |
else: | |
s,e = int(el), int(el) | |
return s,e | |
def ranges_to_indexes(range_string): | |
'''Converts a string containig comma-separated numeric ranges to a list of integers''' | |
idx = [] | |
for x in range_string.split(','): | |
start, end = parse_range_string(x) | |
idx.extend(np.arange(start, end+1)) | |
return np.array(idx) | |
def parse_contigs(contig_input, pdb_id): | |
''' | |
Input: contig start/end by pdb chain and residue number as in the pdb file | |
ex - B12-17 | |
Output: corresponding start/end indices of the "features" numpy array (idx0) | |
''' | |
contigs = [] | |
for con in contig_input.split(','): | |
pdb_ch = con[0] | |
pdb_s, pdb_e = parse_range_string(con[1:]) | |
np_s = pdb_id.index((pdb_ch, pdb_s)) | |
np_e = pdb_id.index((pdb_ch, pdb_e)) | |
contigs.append([np_s, np_e]) | |
return contigs | |
def mk_feat_hal_and_mappings(hal_2_ref_idx0, pdb_out): | |
##################################### | |
# rearrange ref features according to hal_2_ref_idx0 | |
##################################### | |
#1. find corresponding idx0 in hal and ref | |
hal_idx0 = [] | |
ref_idx0 = [] | |
for hal, ref in enumerate(hal_2_ref_idx0): | |
if ref is not None: | |
hal_idx0.append(hal) | |
ref_idx0.append(ref) | |
hal_idx0 = np.array(hal_idx0, dtype=int) | |
ref_idx0 = np.array(ref_idx0, dtype=int) | |
#2. rearrange the 6D features | |
hal_len = len(hal_2_ref_idx0) | |
if 'feat' in pdb_out: | |
d_feat = pdb_out['feat'].shape[3:] | |
feat_hal = np.zeros((1, hal_len, hal_len) + d_feat) | |
feat_ref = pdb_out['feat'] # (B,L,L,...) | |
feat_hal[:, hal_idx0[:,None], hal_idx0[None,:]] = feat_ref[:, ref_idx0[:,None], ref_idx0[None,:]] | |
else: | |
feat_hal = None | |
#3. make the 1d binary mask, for backwards compatibility | |
hal_2_ref_idx0 = np.array(hal_2_ref_idx0, dtype=np.float32) # convert None to NaN | |
mask_1d = (~np.isnan(hal_2_ref_idx0)).astype(float) | |
mask_1d = mask_1d[None] | |
##################################### | |
# mappings between hal and ref | |
##################################### | |
mappings = { | |
'con_hal_idx0': hal_idx0.tolist(), | |
'con_ref_idx0': ref_idx0.tolist(), | |
'con_hal_pdb_idx': [('A',i+1) for i in hal_idx0], | |
'con_ref_pdb_idx': [pdb_out['pdb_idx'][i] for i in ref_idx0], | |
'mask_1d': mask_1d, | |
} | |
return feat_hal, mappings | |
def scatter_feats(template_mask, feat_1d_ref=None, feat_2d_ref=None, pdb_idx=None): | |
''' | |
Scatters 1D and/or 2D reference features according to mappings in hal_2_ref_idx0 | |
Inputs | |
---------- | |
hal_2_ref_idx0: (list; length=L_hal) | |
List mapping hal_idx0 positions to ref_idx0 positions. | |
"None" used for indices that do not map to ref. | |
ex: [None, None, 3, 4, 5, None, None, None, 34, 35, 36] | |
feat_1d_ref: (np.array; (batch, L_ref, ...)) | |
1D refence features to scatter | |
feat_1d_ref: (np.array; (batch, L_ref, L_ref, ...)) | |
pdb_idx: (list) | |
List of pdb chain and residue numbers, in the order that pdb features were read/parsed. | |
Outputs | |
---------- | |
feat_1d_hal: (np.array, (batch, L_hal, ...)) | |
Scattered 1d reference features. "None" mappings are 0. | |
feat_2d_hal: (np.array, (batch, L_hal, L_hal, ...)) | |
Scattered 2d reference features. "None" mappings are 0. | |
mappings: (dict) | |
Keeps track of corresponding possitions in ref and hal proteins. | |
''' | |
hal_2_ref_idx0, _ = contigs.sample_mask(template_mask, pdb_idx) | |
out = {} | |
# Find corresponding idx0 in hal and ref | |
hal_idx0 = [] | |
ref_idx0 = [] | |
hal_len = len(hal_2_ref_idx0) | |
for hal, ref in enumerate(hal_2_ref_idx0): | |
if ref is not None: | |
hal_idx0.append(hal) | |
ref_idx0.append(ref) | |
hal_idx0 = np.array(hal_idx0, dtype=int) | |
ref_idx0 = np.array(ref_idx0, dtype=int) | |
# Make the 1d binary mask, for backwards compatibility | |
hal_2_ref_idx0 = np.array(hal_2_ref_idx0, dtype=np.float32) # convert None to NaN | |
mask_1d = (~np.isnan(hal_2_ref_idx0)).astype(float) | |
mask_1d = mask_1d[None] | |
# scatter 2D features | |
if feat_2d_ref is not None: | |
B = feat_2d_ref.shape[0] | |
d_feat = feat_2d_ref.shape[3:] | |
feat_2d_hal = np.zeros((B, hal_len, hal_len)+d_feat) | |
feat_2d_hal[:, hal_idx0[:,None], hal_idx0[None,:]] = feat_2d_ref[:, ref_idx0[:,None], ref_idx0[None,:]] | |
out['feat_2d_hal'] = feat_2d_hal | |
# scatter 1D features | |
if feat_1d_ref is not None: | |
B = feat_1d_ref.shape[0] | |
d_feat = feat_1d_ref.shape[2:] | |
feat_1d_hal = np.zeros((B, hal_len)+d_feat) | |
feat_1d_hal[:, hal_idx0] = feat_1d_ref[:, ref_idx0] | |
out['feat_1d_hal'] = feat_1d_hal | |
# Mappings between hal and ref | |
mappings = { | |
'con_hal_idx0': hal_idx0.tolist(), | |
'con_ref_idx0': ref_idx0.tolist(), | |
'mask_1d': mask_1d, | |
} | |
if pdb_idx is not None: | |
mappings.update({ | |
'con_hal_pdb_idx': [('A',i+1) for i in hal_idx0], | |
'con_ref_pdb_idx': [pdb_idx[i] for i in ref_idx0], | |
}) | |
out['mappings'] = mappings | |
return out | |
def scatter_contigs(contigs, pdb_out, L_range, keep_order=False, min_gap=0): | |
''' | |
Randomly places contigs in a protein within the length range. | |
Inputs | |
Contig: A continuous range of residues from the pdb. | |
Inclusive of the begining and end | |
Must start with the chain number. Comma separated | |
ex: B6-11,A12-19 | |
pdb_out: dictionary from the prep_input function | |
L_range: String range of possible lengths. | |
ex: 90-110 | |
ex: 70 | |
keep_order: keep contigs in the provided order or randomly permute | |
min_gap: minimum number of amino acids separating contigs | |
Outputs | |
feat_hal: target pdb features to hallucinate | |
mappings: dictionary of ways to convert from the hallucinated protein | |
to the reference protein | |
''' | |
ref_pdb_2_idx0 = {pdb_idx:i for i, pdb_idx in enumerate(pdb_out['pdb_idx'])} | |
##################################### | |
# make a map from hal_idx0 to ref_idx0. Has None for gap regions | |
##################################### | |
#1. Permute contig order | |
contigs = contigs.split(',') | |
if not keep_order: | |
random.shuffle(contigs) | |
#2. convert to ref_idx0 | |
contigs_ref_idx0 = [] | |
for con in contigs: | |
chain = con[0] | |
s, e = parse_range_string(con[1:]) | |
contigs_ref_idx0.append( [ref_pdb_2_idx0[(chain, i)] for i in range(s, e+1)] ) | |
#3. Add minimum gap size | |
for i in range(len(contigs_ref_idx0) - 1): | |
contigs_ref_idx0[i] += [None] * min_gap | |
#4. Sample protein length | |
L_low, L_high = parse_range_string(L_range) | |
L_hal = np.random.randint(L_low, L_high+1) | |
L_con = 0 | |
for con in contigs_ref_idx0: | |
L_con += len(con) | |
L_gaps = L_hal - L_con | |
if L_gaps <= 1: | |
print("Error: The protein isn't long enough to incorporate all the contigs." | |
"Consider reduce the min_gap or increasing L_range") | |
return | |
#5. Randomly insert contigs into gaps | |
hal_2_ref_idx0 = np.array([None] * L_gaps, dtype=float) # inserting contigs into this | |
n_contigs = len(contigs_ref_idx0) | |
insertion_idxs = np.random.randint(L_gaps + 1, size=n_contigs) | |
insertion_idxs.sort() | |
for idx, con in zip(insertion_idxs[::-1], contigs_ref_idx0[::-1]): | |
hal_2_ref_idx0 = np.insert(hal_2_ref_idx0, idx, con) | |
#6. Convert mask to feat_hal and mappings | |
hal_2_ref_idx0 = [int(el) if ~np.isnan(el) else None for el in hal_2_ref_idx0] # convert nan to None | |
feat_hal, mappings = mk_feat_hal_and_mappings(hal_2_ref_idx0, pdb_out) | |
#7. Generate str of the sampled mask | |
contig_positive = np.array(hal_2_ref_idx0) != None | |
boundaries = np.where(np.diff(contig_positive))[0] | |
start_idx0 = np.concatenate([np.array([0]), boundaries+1]) | |
end_idx0 = np.concatenate([boundaries, np.array([contig_positive.shape[0]])-1]) | |
lengths = end_idx0 - start_idx0 + 1 | |
is_contig = contig_positive[start_idx0] | |
sampled_mask = [] | |
con_counter = 0 | |
for i, is_con in enumerate(is_contig): | |
if is_con: | |
sampled_mask.append(contigs[con_counter]) | |
con_counter += 1 | |
else: | |
len_gap = lengths[i] | |
sampled_mask.append(f'{len_gap}-{len_gap}') | |
sampled_mask = ','.join(sampled_mask) | |
mappings['sampled_mask'] = sampled_mask | |
return feat_hal, mappings | |
def get_receptor_contig(ref_pdb_idx): | |
rec_pdb_idx = [idx for idx in ref_pdb_idx if idx[0]=='R'] | |
return SampledMask.contract(rec_pdb_idx) | |
def mk_con_to_set(mask, set_id=None, args=None, ref_pdb_idx=None): | |
''' | |
Maps a mask or list of contigs to a set_id. If no set_id is provided, it treats | |
everything as set 0. | |
Input | |
----------- | |
mask (str): Mask or list of contigs. Ex: 3,B6-11,12,A12-19,9 or Ex: B6-11,A12-19 | |
ref_pdb_idx (List(ch, res)): pdb idxs of the reference pdb. Ex: [(A, 2), (A, 3), ...] | |
args: Arguments object. Must have args.receptor | |
set_id (list): List of integers. Length must match contigs in mask. Ex: [0,1] | |
Output | |
----------- | |
con_to_set (dict): Maps str of contig to integer | |
''' | |
# Extract contigs | |
cons = [l for l in mask.split(',') if l[0].isalpha()] | |
# Assign all contigs to set 0 if set_id is not passed | |
if set_id is None: | |
set_id = [0] * len(cons) | |
con_to_set = dict(zip(cons, set_id)) | |
# Assign receptor to set 0 | |
if args.receptor: | |
receptor_contig = get_receptor_contig(ref_pdb_idx) | |
con_to_set.update({receptor_contig: 0}) | |
return con_to_set | |
def parse_range(_range): | |
if '-' in _range: | |
s, e = _range.split('-') | |
else: | |
s, e = _range, _range | |
return int(s), int(e) | |
def parse_contig(contig): | |
''' | |
Return the chain, start and end residue in a contig or gap str. | |
Ex: | |
'A4-8' --> 'A', 4, 8 | |
'A5' --> 'A', 5, 5 | |
'4-8' --> None, 4, 8 | |
'A' --> 'A', None, None | |
''' | |
# is contig | |
if contig[0].isalpha(): | |
ch = contig[0] | |
if len(contig) > 1: | |
s, e = parse_range(contig[1:]) | |
else: | |
s, e = None, None | |
# is gap | |
else: | |
ch = None | |
s, e = parse_range(contig) | |
return ch, s, e | |
def mask_as_list(sampled_mask): | |
''' | |
Make a length L_hal list, with each position pointing to a ref_pdb_idx (or None) | |
''' | |
mask_list = [] | |
for l in sampled_mask.split(','): | |
ch, s, e = parse_contig(l) | |
# contig | |
if ch is not None: | |
mask_list += [(ch, idx) for idx in range(s, e+1)] | |
# gap | |
else: | |
mask_list += [None for _ in range(s, e+1)] | |
return mask_list | |
def mask_subset(sampled_mask, subset): | |
''' | |
Returns a 1D boolean array of where a subset of the contig is in the hallucinated protein | |
Input | |
--------- | |
subset (str): Some chain and residue subset of the contigs. Ex: A10-15 | |
Can also just pass chain. All contig residues from that chain are selected. Ex: R | |
Ouput | |
--------- | |
m_1d (np.array): Boolean array where subset appears in the hallucinated protein | |
''' | |
mask_list = mask_as_list(sampled_mask) | |
m_1d = [] | |
ch_subset, s, e = parse_contig(subset) | |
assert ch_subset.isalpha(), '"Subset" must include a chain reference' | |
if (s is None) or (e is None): | |
s = -np.inf | |
e = np.inf | |
for l in mask_list: | |
if l is None: | |
continue | |
ch, idx = l | |
if (ch == ch_subset) and (idx >= s) and (idx <= e): | |
m_1d.append(True) | |
else: | |
m_1d.append(False) | |
return np.array(m_1d) | |
def mk_cce_and_hal_mask_2d(sampled_mask, con_to_set=None): | |
''' | |
Makes masks for ij pixels where the cce and hallucination loss should be applied. | |
Inputs | |
--------------- | |
sampled_mask (str): String of where contigs should be applied. Ex: 3,B6-11,12,A12-19,9 | |
cce_cutoff (float): Apply cce loss to cb-cb distances less than this value. Angstroms. | |
con_to_set (dict): Dictionary mapping the string of a contig (ex: 'B6-11') to an integer. | |
L_rec (int): Length of the receptor, if hallucinating in the context of the receptor. | |
Outputs | |
--------------- | |
mask_cce (np.array, (L_hal, L_hal)): Boolean array. True where cce loss should be applied. | |
mask_hal (np.array, (L_hal, L_hal)): Boolean array. True where hallucination loss should be applied. | |
''' | |
if con_to_set is None: | |
con_to_set = mk_con_to_set(sampled_mask) | |
# Length of hallucinated protein | |
L_hal, L_max = mask_len(sampled_mask) | |
assert L_hal == L_max, 'A sampled mask must have gaps of a single length.' | |
# Map each contig to a 1D boolean mask | |
m_con = dict() | |
start_idx = 0 | |
for l in sampled_mask.split(','): | |
if l[0].isalpha(): | |
s, e = parse_range_string(l[1:]) | |
L_con = e - s + 1 | |
m = np.zeros(L_hal, dtype=bool) | |
m[start_idx:start_idx+L_con] = True | |
m_con[l] = m | |
start_idx += L_con | |
else: | |
L_gap, _ = parse_range_string(l) | |
start_idx += L_gap | |
# Combine contigs masks from each set to make 2D mask | |
mask_cce = np.zeros((L_hal, L_hal), dtype=bool) | |
for set_id in set(con_to_set.values()): | |
# gather all masks from contigs in the same set | |
masks = [m_con[k] for k,v in con_to_set.items() if v == set_id] | |
mask_1D = np.any(masks, axis=0) | |
update = mask_1D[:,None] * mask_1D[None,:] | |
mask_cce = np.any([mask_cce, update], axis=0) | |
# Make mask_hal | |
mask_hal = ~mask_cce | |
# Don't apply ANY losses on diagonal | |
mask_cce[np.arange(L_hal), np.arange(L_hal)] = False | |
mask_hal[np.arange(L_hal), np.arange(L_hal)] = False | |
# Don't apply ANY losses to receptor | |
m_1d_rec = mask_subset(sampled_mask, 'R') | |
m_2d_rec = m_1d_rec[:, None] * m_1d_rec[None, :] | |
mask_cce *= ~m_2d_rec | |
mask_hal *= ~m_2d_rec | |
return mask_cce, mask_hal | |
def apply_mask(mask, pdb_out): | |
''' | |
Uniformly samples gap lengths, then gathers the ref features | |
into the target hal features | |
Inputs | |
-------------- | |
mask: specify the order and ranges of contigs and gaps | |
Contig - A continuous range of residues from the pdb. | |
Inclusive of the begining and end | |
Must start with the chain number | |
ex: B6-11 | |
Gap - a gap length or a range of gaps lengths the | |
model is free to hallucinate | |
Gap ranges are inclusive of the end | |
ex: 9-21 | |
ex - '3,B6-11,9-21,A36-42,20-30,A12-24,3-6' | |
pdb_out: dictionary from the prep_input function | |
Outputs | |
------------- | |
feat_hal: features from pdb_out scattered according to the sampled mask | |
mappings: dict keeping track of corresponding positions in the ref and hal features | |
''' | |
ref_pdb_2_idx0 = {pdb_idx:i for i, pdb_idx in enumerate(pdb_out['pdb_idx'])} | |
#1. make a map from hal_idx0 to ref_idx0. Has None for gap regions | |
hal_2_ref_idx0 = [] | |
sampled_mask = [] | |
for el in mask.split(','): | |
if el[0].isalpha(): # el is a contig | |
sampled_mask.append(el) | |
chain = el[0] | |
s,e = parse_range_string(el[1:]) | |
for i in range(s, e+1): | |
idx0 = ref_pdb_2_idx0[(chain, i)] | |
hal_2_ref_idx0.append(idx0) | |
else: # el is a gap | |
# sample gap length | |
s,e = parse_range_string(el) | |
gap_len = np.random.randint(s, e+1) | |
hal_2_ref_idx0 += [None]*gap_len | |
sampled_mask.append(f'{gap_len}-{gap_len}') | |
#2. Convert mask to feat_hal and mappings | |
feat_hal, mappings = mk_feat_hal_and_mappings(hal_2_ref_idx0, pdb_out) | |
#3. Record the mask that was sampled | |
mappings['sampled_mask'] = ','.join(sampled_mask) | |
return feat_hal, mappings | |
def sample_mask(mask, pdb_idx): | |
''' | |
Uniformly samples gap lengths, then gathers the ref features | |
into the target hal features | |
Inputs | |
-------------- | |
mask: specify the order and ranges of contigs and gaps | |
Contig - A continuous range of residues from the pdb. | |
Inclusive of the begining and end | |
Must start with the chain number | |
ex: B6-11 | |
Gap - a gap length or a range of gaps lengths the | |
model is free to hallucinate | |
Gap ranges are inclusive of the end | |
ex: 9-21 | |
ex - '3,B6-11,9-21,A36-42,20-30,A12-24,3-6' | |
Outputs | |
------------- | |
hal_2_ref_idx0: (list; length=L_hal) | |
List mapping hal_idx0 positions to ref_idx0 positions. | |
"None" used for indices that do not map to ref. | |
ex: [None, None, 3, 4, 5, None, None, None, 34, 35, 36] | |
sampled_mask: (str) | |
string of the sampled mask, so the transformations can be reapplied | |
ex - '3-3,B6-11,9-9,A36-42,20-20,A12-24,5-5' | |
''' | |
ref_pdb_2_idx0 = {pdb_i:i for i, pdb_i in enumerate(pdb_idx)} | |
#1. make a map from hal_idx0 to ref_idx0. Has None for gap regions | |
hal_2_ref_idx0 = [] | |
sampled_mask = [] | |
for el in mask.split(','): | |
if el[0].isalpha(): # el is a contig | |
sampled_mask.append(el) | |
chain = el[0] | |
s,e = parse_range_string(el[1:]) | |
for i in range(s, e+1): | |
idx0 = ref_pdb_2_idx0[(chain, i)] | |
hal_2_ref_idx0.append(idx0) | |
else: # el is a gap | |
# sample gap length | |
s,e = parse_range_string(el) | |
gap_len = np.random.randint(s, e+1) | |
hal_2_ref_idx0 += [None]*gap_len | |
sampled_mask.append(f'{gap_len}-{gap_len}') | |
return hal_2_ref_idx0, sampled_mask | |
class GapResampler(): | |
def __init__(self, use_bkg=True): | |
''' | |
''' | |
self.counts_passed = {} # dictionary for tallying counts of gap lengths for designs passing some threshold | |
self.counts_bkg = {} | |
self.use_bkg = use_bkg | |
def clean_mask(self, mask): | |
''' | |
Makes mask into a cononical form. | |
Ensures masks always alternate gap, contig and that | |
masks begin and end with a gap (even of length 0) | |
Input | |
----------- | |
masks: list of masks (str). Mask format: comma separted list | |
of alternating gap_length (int or int-int), contig. | |
Ex - 9,A12-19,15,B45-52 OR 9-9,A12-19,15-15,B45-52 | |
Output | |
----------- | |
A canonicalized mask. Ex: N,9,A12-19,15,B45-52,0,C | |
''' | |
mask = mask.split(',') | |
mask_out = [] | |
was_contig = True | |
was_gap = False | |
for i, el in enumerate(mask): | |
is_contig = el[0].isalpha() | |
is_gap = not is_contig | |
is_last = i == len(mask) - 1 | |
# accepting gaps as either x-x or just x | |
if is_gap: | |
if '-' in el: | |
x1, x2 = el.split('-') | |
if x1 != x2: | |
print(f"Error: Gap must not be a range: {mask}") | |
return None | |
gap = x1 | |
else: | |
gap = el | |
if is_contig: | |
contig = el | |
# gap -> contig: just append new contig | |
if (was_gap and is_contig): | |
mask_out.append(contig) | |
# contig -> gap: just append gap | |
elif (was_contig and is_gap): | |
mask_out.append(gap) | |
# contig -> contig: insert gap of 0, then add contig | |
elif (was_contig and is_contig): | |
mask_out.append('0') | |
mask_out.append(contig) | |
# gap -> gap: add them | |
elif (was_gap and is_gap): | |
combined_len = int(mask_out[-1]) + int(gap) | |
mask_out[-1] = str(combined_len) | |
# ensure last mask element is a gap | |
if (is_last and is_contig): | |
mask_out.append('0') | |
# update what previous element was | |
was_contig = el[0].isalpha() | |
was_gap = ~is_contig | |
# add 'N' and 'C' contigs | |
mask_out.insert(0, 'N') | |
mask_out.append('C') | |
return ','.join(mask_out) | |
def add_mask(self, mask, counting_dict): | |
''' | |
Adds counts of gap lengths to counting_dict | |
Inputs | |
----------- | |
masks: list of masks (str). Mask format: comma separted list | |
of alternating gap_length (int or int-int), contig. | |
Ex - 9,A12-19,15,B45-52 OR 9-9,A12-19,15-15,B45-52 | |
''' | |
mask = self.clean_mask(mask) | |
mask = mask.split(',') | |
n_gaps = len(mask) // 2 | |
# count occurances of contig,gap,contig triples | |
for i in range(n_gaps): | |
con1, gap, con2 = mask[2*i : 2*i+3] | |
# count gap length | |
if con1 in counting_dict: | |
if (gap, con2) in counting_dict[con1]: | |
counting_dict[con1][(gap, con2)] += 1 | |
else: | |
counting_dict[con1][(gap, con2)] = 1 | |
else: | |
counting_dict[con1] = {(gap, con2): 1} | |
def add_mask_pass(self, mask): | |
''' | |
Add a mask that passed to self.counts_passed | |
''' | |
self.add_mask(mask, self.counts_passed) | |
def add_mask_bkg(self, mask): | |
''' | |
Add a mask that passed to self.counts_bkg | |
''' | |
self.add_mask(mask, self.counts_bkg) | |
def get_enrichment(self): | |
''' | |
Calculate the ratio of counts_passed / count_bkg | |
Also notes all contigs | |
''' | |
if self.use_bkg is False: | |
print('Please pass in background masks and set self.use_bkg=True') | |
return | |
self.counts_enrich = copy.copy(self.counts_passed) | |
self.con_all = set() | |
for con1 in self.counts_enrich.keys(): | |
self.con_all |= set([con1]) | |
for gap, con2 in self.counts_enrich[con1].keys(): | |
self.con_all |= set([con2]) | |
bkg = self.counts_bkg[con1][(gap, con2)] | |
cnt = self.counts_passed[con1][(gap, con2)] | |
self.counts_enrich[con1][(gap, con2)] = cnt / bkg | |
def sample_mask(self): | |
''' | |
Sample a mask | |
''' | |
searching = True | |
while searching: | |
n_gaps = len(self.con_all) - 1 | |
mask = ['N'] | |
if self.use_bkg: | |
counts = self.counts_enrich | |
else: | |
counts = self.counts_passed | |
for i in range(n_gaps): | |
con_last = mask[-1] | |
# only allow jump to C as last option | |
if i == n_gaps - 1: | |
con_used = set(mask[::2]) | |
else: | |
con_used = set(mask[::2]+['C']) | |
con_free = self.con_all - con_used | |
# get available "jumps" (con -> gap, con) you can make | |
jumps_all = counts[con_last] | |
jumps_free = {k:v for k,v in jumps_all.items() if k[1] in con_free} | |
if len(jumps_free) == 0: | |
print('No available jumps to continue the mask. Sampling again...') | |
else: | |
# normalize counts and sample move | |
mvs, cnt = zip(*jumps_free.items()) | |
cnt = np.array(cnt) | |
prob = cnt / cnt.sum() | |
idx = np.random.choice(len(prob), p=prob) | |
mv = mvs[idx] | |
# add to the mask | |
mask.append(mv[0]) | |
mask.append(mv[1]) | |
# check that mask has the right number of elements | |
if len(mask) == 2*n_gaps + 1: | |
searching = False | |
else: | |
searching = True | |
return ','.join(mask[1:-1]) | |
def gaps_as_ranges(self, mask): | |
''' | |
Convert gaps of a single int to ranges, for | |
backwards compatibility reasons | |
''' | |
mask_out = [] | |
for el in mask.split(','): | |
if el[0].isalpha(): | |
mask_out.append(el) | |
else: | |
mask_out.append(f'{el}-{el}') | |
return ','.join(mask_out) | |
def recover_mask(trb): | |
''' | |
Recover the string of the sampled mask given the trb file | |
''' | |
L_hal = trb['mask_contig'].shape[0] | |
mask = [] | |
for idx0 in range(L_hal): | |
# what is the current idx | |
if idx0 in trb['con_hal_idx0']: | |
is_con = True | |
is_gap = False | |
else: | |
is_con = False | |
is_gap = True | |
# dealing with the first entry | |
if idx0 == 0: | |
if is_gap: | |
L_gap = 1 | |
elif is_con: | |
ch, idx = trb['con_ref_pdb_idx'][ trb['con_hal_idx0'].tolist().index(idx0) ] | |
con_start = f'{ch}{idx}' | |
# take action based on what happend last time | |
else: | |
if (was_gap) and (is_gap): | |
L_gap +=1 | |
#elif (was_con) and (is_con): | |
# continue | |
elif (was_gap) and (is_con): | |
# end gap | |
mask.append(f'{L_gap}-{L_gap}') | |
# start con | |
ch, idx = trb['con_ref_pdb_idx'][ trb['con_hal_idx0'].tolist().index(idx0) ] | |
con_start = f'{ch}{idx}' | |
elif (was_con) and (is_gap): | |
# end con | |
ch, idx = trb['con_ref_pdb_idx'][ trb['con_hal_idx0'].tolist().index(idx0) ] | |
mask.append(f'{con_start}-{idx}') | |
# start gap | |
L_gap = 1 | |
# dealing with last entry | |
if idx0 == L_hal-1: | |
if is_gap: | |
mask.append(f'{L_gap}-{L_gap}') | |
elif is_con: # (edge case not handled: con starts and ends on last idx) | |
ch, idx = trb['con_ref_pdb_idx'][ trb['con_hal_idx0'].tolist().index(idx0-1) ] | |
mask.append(f'{con_start}-{idx}') | |
# update what last position was | |
was_con = copy.copy(is_con) | |
was_gap = copy.copy(is_gap) | |
return ','.join(mask) | |
def mask_len(mask): | |
''' | |
Calculate the min and max possible length that can | |
be sampled given a mask | |
''' | |
L_min = 0 | |
L_max = 0 | |
for el in mask.split(','): | |
if el[0].isalpha(): # is con | |
con_s, con_e = el[1:].split('-') | |
con_s, con_e = int(con_s), int(con_e) | |
L_con = con_e - con_s + 1 | |
L_min += L_con | |
L_max += L_con | |
else: # is gap | |
if '-' in el: | |
gap_min, gap_max = el.split('-') | |
gap_min, gap_max = int(gap_min), int(gap_max) | |
L_min += gap_min | |
L_max += gap_max | |
else: | |
L_min += int(el) | |
L_max += int(el) | |
return L_min, L_max | |
class SampledMask(): | |
def __init__(self, mask_str, ref_pdb_idx, con_to_set=None): | |
self.str = mask_str | |
self.L_hal = len(self) | |
self.L_ref = len(ref_pdb_idx) | |
################# | |
# con indices in hal and ref | |
################# | |
self.ref_pdb_idx = ref_pdb_idx | |
self.hal_pdb_idx = [('A', i) for i in range(1, len(self)+1)] | |
hal_idx0 = 0 | |
con_ref_pdb_idx = [] | |
con_hal_pdb_idx = [] | |
con_ref_idx0 = [] | |
con_hal_idx0 = [] | |
for l in mask_str.split(','): | |
ch, s, e = SampledMask.parse_contig(l) | |
# contig | |
if ch: | |
for res in range(s, e+1): | |
con_ref_pdb_idx.append((ch, res)) | |
con_hal_pdb_idx.append(('A', hal_idx0+1)) | |
con_ref_idx0.append(self.ref_pdb_idx.index((ch, res))) | |
con_hal_idx0.append(hal_idx0) | |
hal_idx0 += 1 | |
# gap | |
else: | |
for _ in range(s): | |
hal_idx0 += 1 | |
self.con_mappings = { | |
'ref_pdb_idx': con_ref_pdb_idx, | |
'hal_pdb_idx': con_hal_pdb_idx, | |
'ref_idx0': con_ref_idx0, | |
'hal_idx0': con_hal_idx0, | |
} | |
################# | |
# con_to_set mapping | |
################# | |
if con_to_set: | |
self.con_to_set = con_to_set | |
else: | |
contigs = self.get_contigs() | |
self.con_to_set = dict(zip(contigs, len(contigs)*[0])) | |
# set_to_con mapping | |
set_to_con = {} | |
for k, v in self.con_to_set.items(): | |
set_to_con[v] = set_to_con.get(v, []) + [k] # invert a dictionary with non-unique values | |
self.set_to_con = set_to_con | |
def __len__(self,): | |
_, L_max = self.mask_len(self.str) | |
return L_max | |
def map(self, sel, src, dst): | |
''' | |
Convert the contig selection in one indexing scheme to another. | |
Will return None if selection is not in a contig. | |
Input | |
---------- | |
sel (str): selection of a contig range or idx0 range. Can take multiple comma separated values of same type. Ex: A5-10,B2-8 or 3-8,14-21 | |
src (str): <'ref', 'hal'> | |
dst (str): <'ref_pdb_idx', 'hal_pdb_idx', 'ref_idx0', 'hal_idx0> | |
''' | |
out = [] | |
for con in sel.split(','): | |
ch, s, e = SampledMask.parse_contig(con) | |
# selection type is pdb_idx | |
if ch: | |
src_long = f'{src}_pdb_idx' | |
mapping = dict(zip(self.con_mappings[src_long], self.con_mappings[dst])) | |
out += [mapping.get((ch, res)) for res in range(s, e+1)] | |
# selection type is idx0 | |
else: | |
src_long = f'{src}_idx0' | |
mapping = dict(zip(self.con_mappings[src_long], self.con_mappings[dst])) | |
out += [mapping.get(i) for i in range(s, e+1)] | |
return out | |
def expand(mask_str): | |
''' | |
Ex: '2,A3-5,3' --> [None, None, (A,3), (A,4), (A,5), None, None, None] | |
''' | |
expanded = [] | |
for l in mask_str.split(','): | |
ch, s, e = SampledMask.parse_contig(l) | |
# contig | |
if ch: | |
expanded += [(ch, res) for res in range(s, e+1)] | |
# gap | |
else: | |
expanded += [None for _ in range(s)] | |
return expanded | |
def contract(pdb_idx): | |
''' | |
Inverse of expand | |
Ex: [None, None, (A,3), (A,4), (A,5), None, None, None] --> '2,A3-5,3' | |
''' | |
contracted = [] | |
l_prev = (None, -200) | |
first_el_written = False | |
for l_curr in pdb_idx: | |
if l_curr is None: | |
l_curr = (None, -100) | |
# extend gap | |
if l_curr == l_prev: | |
L_gap += 1 | |
# extend con | |
elif l_curr == (l_prev[0], l_prev[1]+1): | |
con_e = l_curr[1] | |
# new gap | |
elif (l_curr != l_prev) and (l_curr[0] is None): | |
# write prev con | |
if 'con_ch' in locals(): | |
contracted.append(f'{con_ch}{con_s}-{con_e}') | |
L_gap = 1 | |
# new con | |
elif (l_curr != l_prev) and isinstance(l_curr[0], str): | |
# write prev con | |
if isinstance(l_prev[0], str) and ('con_ch' in locals()): | |
contracted.append(f'{con_ch}{con_s}-{con_e}') | |
# write prev gap | |
elif 'L_gap' in locals(): | |
contracted.append(str(L_gap)) | |
con_ch = l_curr[0] | |
con_s = l_curr[1] | |
con_e = l_curr[1] | |
# update l_prev | |
l_prev = l_curr | |
# write last element | |
if isinstance(l_prev[0], str) and ('con_ch' in locals()): | |
contracted.append(f'{con_ch}{con_s}-{con_e}') | |
elif 'L_gap' in locals(): | |
contracted.append(str(L_gap)) | |
return ','.join(contracted) | |
def subset(self, sub): | |
''' | |
Make a mask_str that is a subset of the original mask_str | |
Ex: self.mask_str = '2,A5-20,4', sub='A5-10' --> '2,A5-10,14' | |
''' | |
# map from hal_idx0 to ref_pdb_idx | |
hal_idx0 = self.map(sub, 'ref', 'hal_idx0') | |
ref_pdb_idx = SampledMask.expand(sub) | |
mapping = dict(zip(hal_idx0, ref_pdb_idx)) | |
expanded = [mapping.get(idx0) for idx0 in range(len(self))] | |
return self.contract(expanded) | |
def mask_len(self, mask): | |
''' | |
Technically, can take both sampled and unsampled mask | |
''' | |
L_min = 0 | |
L_max = 0 | |
for l in self.str.split(','): | |
ch, s, e = SampledMask.parse_contig(l) | |
# contig | |
if ch: | |
L_min += e - s + 1 | |
L_max += e - s + 1 | |
# gap | |
else: | |
L_min += s | |
L_max += e | |
return L_min, L_max | |
def get_contigs(self, include_receptor=True): | |
''' | |
Get a list of all contigs in the mask | |
''' | |
[con for con in self.str.split(',') if SampledMask.parse_contig(con)[0]] | |
contigs = [] | |
for con in self.str.split(','): | |
ch = SampledMask.parse_contig(con)[0] | |
if ch == 'R' and include_receptor == False: | |
continue | |
if ch: | |
contigs.append(con) | |
return contigs | |
def get_gaps(self,): | |
''' | |
Get a list of all gaps in the mask | |
''' | |
return [con for con in self.str.split(',') if SampledMask.parse_contig(con)[0] is None] | |
def parse_range(_range): | |
if '-' in _range: | |
s, e = _range.split('-') | |
else: | |
s, e = _range, _range | |
return int(s), int(e) | |
def parse_contig(contig): | |
''' | |
Return the chain, start and end residue in a contig or gap str. | |
Ex: | |
'A4-8' --> 'A', 4, 8 | |
'A5' --> 'A', 5, 5 | |
'4-8' --> None, 4, 8 | |
'A' --> 'A', None, None | |
''' | |
# is contig | |
if contig[0].isalpha(): | |
ch = contig[0] | |
if len(contig) > 1: | |
s, e = SampledMask.parse_range(contig[1:]) | |
else: | |
s, e = None, None | |
# is gap | |
else: | |
ch = None | |
s, e = SampledMask.parse_range(contig) | |
return ch, s, e | |
def remove_diag(self, m_2d): | |
''' | |
Set the diagonal of a 2D boolean array to False | |
''' | |
L = m_2d.shape[0] | |
m_2d[np.arange(L), np.arange(L)] = False | |
return m_2d | |
def get_receptor_contig(self,): | |
''' | |
Returns None if there is no chain R in the mask_str | |
''' | |
receptor_contig = [l for l in self.get_contigs() if 'R' in l] | |
if len(receptor_contig) == 0: | |
receptor_contig = None | |
else: | |
receptor_contig = ','.join(receptor_contig) | |
return receptor_contig | |
def remove_receptor(self, m_2d): | |
''' | |
Remove intra-receptor contacts (chain R) from a mask | |
''' | |
receptor_contig = self.get_receptor_contig() | |
if receptor_contig: # has chain R | |
m_1d = np.zeros(self.L_hal, dtype=bool) | |
idx = np.array(self.map(receptor_contig, 'ref', 'hal_idx0')) | |
m_1d[idx] = True | |
update = m_1d[:, None] * m_1d[None, :] | |
m_2d = m_2d * ~update | |
return m_2d | |
def get_mask_con(self, include_receptor=False): | |
# Make a 2D boolean mask for each contig set | |
L = self.L_hal | |
mask_con = np.zeros([L, L], dtype=bool) | |
for set_id, contigs in self.set_to_con.items(): | |
m_1d = np.zeros(L, dtype=bool) | |
for con in contigs: | |
idx = self.map(con, 'ref', 'hal_idx0') | |
idx = [l for l in idx if l != None] | |
idx = np.array(idx, dtype=int) | |
m_1d[idx] = True | |
update = m_1d[:, None] * m_1d[None, :] | |
mask_con = np.any([mask_con, update], axis=0) | |
# clean up | |
mask_con = self.remove_diag(mask_con) | |
if not include_receptor: | |
mask_con = self.remove_receptor(mask_con) | |
return mask_con | |
def get_mask_hal(self,): | |
mask_hal = ~self.get_mask_con() | |
mask_hal = self.remove_diag(mask_hal) | |
mask_hal = self.remove_receptor(mask_hal) | |
return mask_hal | |
def get_mask_cce(self, pdb, cce_cutoff=20., include_receptor=False): | |
''' | |
Remove ij pixels where contig distances are greater than cce_cutoff. | |
''' | |
# start with mask_con | |
mask_con = self.get_mask_con(include_receptor=include_receptor) | |
# get ref dists | |
xyz_ref = torch.tensor(pdb['xyz'][:,:3,:]).float() | |
c6d_ref = geometry.xyz_to_c6d(xyz_ref[None].permute(0,2,1,3),{'DMAX':20.0}).numpy() | |
dist = c6d_ref[0,:,:,0] # (L_ref, L_ref) | |
# scatter | |
dist_scattered = self.scatter_2d(dist) | |
# apply cce cuttoff | |
update = dist_scattered < cce_cutoff | |
mask_cce = np.all([mask_con, update], axis=0) | |
return mask_cce | |
def scatter_2d(self, ref_feat_2d): | |
''' | |
Inputs | |
--------- | |
ref_feat_2d (np.array; (L_ref, L_ref, ...)): Features to be scattered. The first two leading dimensions must be equal to L_ref. | |
''' | |
assert ref_feat_2d.shape[:2] == (self.L_ref, self.L_ref), 'ERROR: feat_2d must have leading dimensions of (L_ref, L_ref)' | |
trailing_dims = ref_feat_2d.shape[2:] | |
dtype = ref_feat_2d.dtype | |
hal_feat_2d = np.zeros((self.L_hal, self.L_hal)+trailing_dims, dtype=dtype) | |
con_hal_idx0 = np.array(self.con_mappings['hal_idx0']) | |
ref_hal_idx0 = np.array(self.con_mappings['ref_idx0']) | |
hal_feat_2d[con_hal_idx0[:, None], con_hal_idx0[None, :]] = ref_feat_2d[ref_hal_idx0[:, None], ref_hal_idx0[None, :]] | |
return hal_feat_2d | |
def scatter_1d(self, ref_feat_1d): | |
''' | |
Inputs | |
--------- | |
ref_feat_1d (np.array; (L_ref, ...)): Features to be scattered. The first leading dimension must be equal to L_ref. | |
''' | |
assert ref_feat_1d.shape[0] == self.L_ref, 'ERROR: feat_1d must have leading dimensions of (L_ref,)' | |
trailing_dims = ref_feat_1d.shape[1:] | |
dtype = ref_feat_1d.dtype | |
hal_feat_1d = np.zeros((self.L_hal,)+trailing_dims, dtype=dtype) | |
con_hal_idx0 = np.array(self.con_mappings['hal_idx0']) | |
ref_hal_idx0 = np.array(self.con_mappings['ref_idx0']) | |
hal_feat_1d[con_hal_idx0] = ref_feat_1d[ref_hal_idx0] | |
return hal_feat_1d | |
def idx_for_template(self, gap=200): | |
''' | |
Essentially return hal_idx0, except have a large jump for chain B, | |
to simulate a chain break. If B contains internal jumps in residue | |
numbering, these are preserved. | |
''' | |
is_rec = self.m1d_receptor() | |
resi_rec = np.array([idx[1] for idx in SampledMask.expand(self.str) | |
if idx is not None and idx[0]=='R']) | |
L_binder = sum(~is_rec) | |
if len(resi_rec)>0: | |
if is_rec[0]: | |
# receptor first | |
idx_tmpl = np.arange(resi_rec[-1]+gap+1, resi_rec[-1]+gap+1+L_binder) | |
idx_tmpl = np.concatenate([resi_rec, idx_tmpl]) | |
else: | |
# receptor second | |
idx_tmpl = np.arange(L_binder) | |
if resi_rec[0] <= idx_tmpl[-1]+gap: | |
resi_rec += idx_tmpl[-1] - resi_rec[0] + gap + 1 | |
idx_tmpl = np.concatenate([idx_tmpl, resi_rec]) | |
else: | |
#when no receptor | |
idx_tmpl = np.arange(L_binder) | |
return idx_tmpl | |
def m1d_receptor(self,): | |
''' | |
Get a boolean array, True if the position corresponds to the receptor | |
''' | |
m1d = [(l is not None) and (l[0] == 'R') for l in SampledMask.expand(self.str)] | |
return np.array(m1d) | |
def erode(self, N_term=True, C_term=True): | |
''' | |
Reduce non-receptor contigs by 1 residue from the N and/or C terminus. | |
''' | |
x = SampledMask.expand(self.str) | |
if N_term: | |
for i, l in enumerate(x): | |
if (l is not None) and (l[0] != 'R'): | |
x[i] = None | |
break | |
if C_term: | |
x = x[::-1] | |
for i, l in enumerate(x): | |
if (l is not None) and (l[0] != 'R'): | |
x[i] = None | |
break | |
x = x[::-1] | |
self.str = self.contract(x) | |
return | |
def len_contigs(self, include_receptor=False): | |
con_str = ','.join(self.get_contigs(include_receptor)) | |
return len(SampledMask.expand(con_str)) | |
def make_template_features(pdb, args, device, hal_2_ref_idx0=None, sm_loss=None): | |
''' | |
Inputs | |
---------- | |
sm_loss: Instance of a contig.SampledMask object used for making the loss masks. | |
''' | |
PARAMS = { | |
"DMIN" : 2.0, | |
"DMAX" : 20.0, | |
"DBINS" : 36, | |
"ABINS" : 36, | |
} | |
if args.use_template: | |
B,T = 1,1 # batch, templates | |
# spoof reference features | |
xyz_t = torch.tensor(pdb['xyz'][:, :3][None, None]) # (batch,templ,nres,3,3) | |
t0d = torch.ones((1,1,3)) # (batch, templ, 3) | |
t2d_ref = xyz_to_t2d(xyz_t=xyz_t, t0d=t0d, params=PARAMS) # (B,T,L,L,...) | |
L_ref = t2d_ref.shape[2] | |
#t1d_ref = torch.ones(size=(B,T,L_ref,3), dtype=torch.float32, device=device) | |
a = 2 * torch.ones([B,T,L_ref], dtype=torch.float32, device=device) | |
b = 0 * torch.ones([B,T,L_ref], dtype=torch.float32, device=device) | |
c = 1 * torch.ones([B,T,L_ref], dtype=torch.float32, device=device) | |
t1d_ref = torch.stack([a,b,c], axis=-1) | |
# Get the mask_str for scattering template features | |
#1. Template mask = sampled mask | |
if (args.use_template.lower() == 't') or (args.use_template.lower() == 'true'): | |
sm_tmpl = sm_loss | |
#2. Template mask is a subset of the sampled mask | |
else: | |
subset_contigs = args.use_template | |
if args.receptor: | |
receptor_contig = sm_loss.get_receptor_contig() | |
subset_contigs = ','.join([subset_contigs, receptor_contig]) | |
mask_str_tmpl = sm_loss.subset(subset_contigs) | |
sm_tmpl = SampledMask(mask_str=mask_str_tmpl, ref_pdb_idx=pdb['pdb_idx']) | |
# scatter template features | |
# make leading dims (L,(L),...) | |
t1d_ref = t1d_ref.permute(2,3,0,1) # (L, ..., B, T) | |
t2d_ref = t2d_ref.permute(2,3,4,0,1) # (L, L, ..., B, T) | |
t1d_tmpl = sm_tmpl.scatter_1d(t1d_ref.cpu().numpy()) | |
t2d_tmpl = sm_tmpl.scatter_2d(t2d_ref.cpu().numpy()) | |
# update t2d_tmpl with mask_con (could update with mask_cce instead?) | |
mask_con = sm_tmpl.get_mask_con(include_receptor=True) | |
t2d_tmpl = (t2d_tmpl.T * mask_con.T).T # trick to broadcast arrays if leading dimensions match | |
t1d_tmpl = torch.tensor(t1d_tmpl, device=device) | |
t2d_tmpl = torch.tensor(t2d_tmpl, device=device) | |
# Permute B and T dims back to front | |
t1d_tmpl = t1d_tmpl.permute(2,3,0,1) | |
t2d_tmpl = t2d_tmpl.permute(3,4,0,1,2) | |
# Make last 3 idx of last dim all 1 to mimick Ivan's template feature | |
t2d_tmpl[..., -3:] = 1. | |
idx = torch.tensor(sm_tmpl.idx_for_template(gap=200), device=device)[None] | |
net_kwargs = { | |
'idx': idx, | |
't1d': t1d_tmpl, | |
't2d': t2d_tmpl | |
} | |
elif args.template_pdbs is not None: | |
B,T = 1, len(args.template_pdbs) # batch, templates | |
# get xyz features of all templates | |
xyz_t = [torch.tensor(parse_pdb(f_pdb)['xyz'][:, :3]) for f_pdb in args.template_pdbs] | |
xyz_t = torch.stack(xyz_t, axis=0)[None] # (batch, template, nres, 3, 3) | |
t0d = torch.ones(B,T,3) | |
t2d_tmpl = xyz_to_t2d(xyz_t=xyz_t, t0d=t0d, params=PARAMS).to(device) # (B,T,L,L,...) | |
L_tmpl = t2d_tmpl.shape[2] | |
t1d_tmpl = torch.ones(size=(B,T,L_tmpl,3), dtype=torch.float32, device=device) | |
# spoof pdb idx | |
idx_tmpl = torch.range(0, L_tmpl-1, dtype=torch.long, device=device)[None] | |
# Net() kwargs | |
net_kwargs = { | |
'idx': idx_tmpl, | |
't1d': t1d_tmpl, | |
't2d': t2d_tmpl | |
} | |
else: | |
net_kwargs = {} | |
return net_kwargs | |