Spaces:
Sleeping
Sleeping
File size: 6,460 Bytes
59a9ccf |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 |
import sys, os
import torch
from icecream import ic
import random
import numpy as np
from kinematics import get_init_xyz
sys.path.append('../')
from utils.calc_dssp import annotate_sse
ic.configureOutput(includeContext=True)
def mask_inputs(seq,
msa_masked,
msa_full,
xyz_t,
t1d,
mask_msa,
input_seq_mask=None,
input_str_mask=None,
input_floating_mask=None,
input_t1dconf_mask=None,
loss_seq_mask=None,
loss_str_mask=None,
loss_str_mask_2d=None,
dssp=False,
hotspots=False,
diffuser=None,
t=None,
freeze_seq_emb=False,
mutate_seq=False,
no_clamp_seq=False,
norm_input=False,
contacts=None,
frac_provide_dssp=0.5,
dssp_mask_percentage=[0,100],
frac_provide_contacts=0.5,
struc_cond=False):
"""
Parameters:
seq (torch.tensor, required): (I,L) integer sequence
msa_masked (torch.tensor, required): (I,N_short,L,48)
msa_full (torch,.tensor, required): (I,N_long,L,25)
xyz_t (torch,tensor): (T,L,27,3) template crds BEFORE they go into get_init_xyz
t1d (torch.tensor, required): (I,L,22) this is the t1d before tacking on the chi angles
str_mask_1D (torch.tensor, required): Shape (L) rank 1 tensor where structure is masked at False positions
seq_mask_1D (torch.tensor, required): Shape (L) rank 1 tensor where seq is masked at False positions
t1d_24: is there an extra dimension to input structure confidence?
diffuser: diffuser class
t: time step
NOTE: in the MSA, the order is 20aa, 1x unknown, 1x mask token. We set the masked region to 22 (masked).
For the t1d, this has 20aa, 1x unkown, and 1x template conf. Here, we set the masked region to 21 (unknown).
This, we think, makes sense, as the template in normal RF training does not perfectly correspond to the MSA.
"""
#ic(input_seq_mask.shape)
#ic(seq.shape)
#ic(msa_masked.shape)
#ic(msa_full.shape)
#ic(t1d.shape)
#ic(xyz_t.shape)
#ic(input_str_mask.shape)
#ic(mask_msa.shape)
###########
seq_mask = input_seq_mask
######################
###sequence diffusion###
######################
str_mask = input_str_mask
x_0 = torch.nn.functional.one_hot(seq[0,...],num_classes=22).float()*2-1
seq_diffused = diffuser.q_sample(x_0,t,mask=seq_mask)
seq_tmp=torch.argmax(seq_diffused,axis=-1).to(device=seq.device)
seq=seq_tmp.repeat(seq.shape[0], 1)
###################
###msa diffusion###
###################
### msa_masked ###
#ic(msa_masked.shape)
B,N,L,_=msa_masked.shape
msa_masked[:,0,:,:22] = seq_diffused
x_0_msa = msa_masked[0,1:,:,:22].float()*2-1
msa_seq_mask = seq_mask.unsqueeze(0).repeat(N-1, 1)
msa_diffused = diffuser.q_sample(x_0_msa,torch.tensor([t]),mask=msa_seq_mask)
msa_masked[:,1:,:,:22] = torch.clone(msa_diffused)
# index 44/45 is insertion/deletion
# index 43 is the masked token NOTE check this
# index 42 is the unknown token
msa_masked[:,0,:,22:44] = seq_diffused
msa_masked[:,1:,:,22:44] = msa_diffused
# insertion/deletion stuff
msa_masked[:,0,~seq_mask,44:46] = 0
### msa_full ###
################
#make msa_full same size as msa_masked
#ic(msa_full.shape)
msa_full = msa_full[:,:msa_masked.shape[1],:,:]
msa_full[:,0,:,:22] = seq_diffused
msa_full[:,1:,:,:22] = msa_diffused
### t1d ###
###########
# NOTE: adjusting t1d last dim (confidence) from sequence mask
t1d = torch.cat((t1d, torch.zeros((t1d.shape[0],t1d.shape[1],1)).float()), -1).to(seq.device)
t1d[:,:,:21] = seq_diffused[...,:21]
#t1d[:,:,21] *= input_t1dconf_mask
#set diffused conf to 0 and everything else to 1
t1d[:,~seq_mask,21] = 0.0
t1d[:,seq_mask,21] = 1.0
t1d[:1,:,22] = 1-t/diffuser.num_timesteps
#to do add structure confidence metric; need to expand dimensions of chkpt b4
#if t1d_24: JG - changed to be default
t1d = torch.cat((t1d, torch.zeros((t1d.shape[0],t1d.shape[1],1)).float()), -1).to(seq.device)
t1d[:,~str_mask,23] = 0.0
t1d[:,str_mask,23] = 1.0
if dssp:
print(f'adding dssp {frac_provide_dssp} of time')
t1d = torch.cat((t1d, torch.zeros((t1d.shape[0],t1d.shape[1],4)).float()), -1).to(seq.device)
#dssp info
#mask some percentage of dssp info in range dssp_mask_percentage[0],dssp_mask_percentage[1]
percentage_mask=random.randint(dssp_mask_percentage[0], dssp_mask_percentage[1])
dssp=annotate_sse(np.array(xyz_t[0,:,1,:].squeeze()), percentage_mask=percentage_mask)
#dssp_unmasked = annotate_sse(np.array(xyz_t[0,:,1,:].squeeze()), percentage_mask=0)
if np.random.rand()>frac_provide_dssp:
print('masking dssp')
dssp[...]=0 #replace with mask token
dssp[:,-1]=1
t1d[...,24:]=dssp
if hotspots:
print(f"adding hotspots {frac_provide_contacts} of time")
t1d = torch.cat((t1d, torch.zeros((t1d.shape[0],t1d.shape[1],1)).float()), -1).to(seq.device)
#mask all contacts some fraction of the time
if np.random.rand()>frac_provide_contacts:
print('masking contacts')
contacts = torch.zeros(L)
t1d[...,-1] = contacts
### xyz_t ###
#############
xyz_t = get_init_xyz(xyz_t[None])
xyz_t = xyz_t[0]
#Sequence masking
xyz_t[:,:,3:,:] = float('nan')
# Structure masking
if struc_cond:
print("non-autoregressive structure conditioning")
r = diffuser.alphas_cumprod[t]
xyz_mask = (torch.rand(xyz_t.shape[1]) > r).to(torch.bool).to(seq.device)
xyz_mask = torch.logical_and(xyz_mask,~str_mask)
xyz_t[:,xyz_mask,:,:] = float('nan')
else:
xyz_t[:,~str_mask,:,:] = float('nan')
### mask_msa ###
################
# NOTE: this is for loss scoring
mask_msa[:,:,~loss_seq_mask] = False
out=dict(
seq= seq,
msa_masked= msa_masked,
msa_full= msa_full,
xyz_t= xyz_t,
t1d= t1d,
mask_msa= mask_msa,
seq_diffused= seq_diffused
)
return out
|