Christina Theodoris
commited on
Commit
•
efec1c4
1
Parent(s):
09276dd
add in silico perturbation module
Browse files- geneformer/__init__.py +4 -0
- geneformer/in_silico_perturber.py +777 -0
- geneformer/in_silico_perturber_stats.py +302 -0
- geneformer/pretrainer.py +2 -1
geneformer/__init__.py
CHANGED
@@ -2,7 +2,11 @@ from . import tokenizer
|
|
2 |
from . import pretrainer
|
3 |
from . import collator_for_cell_classification
|
4 |
from . import collator_for_gene_classification
|
|
|
|
|
5 |
from .tokenizer import TranscriptomeTokenizer
|
6 |
from .pretrainer import GeneformerPretrainer
|
7 |
from .collator_for_gene_classification import DataCollatorForGeneClassification
|
8 |
from .collator_for_cell_classification import DataCollatorForCellClassification
|
|
|
|
|
|
2 |
from . import pretrainer
|
3 |
from . import collator_for_cell_classification
|
4 |
from . import collator_for_gene_classification
|
5 |
+
from . import in_silico_perturber
|
6 |
+
from . import in_silico_perturber_stats
|
7 |
from .tokenizer import TranscriptomeTokenizer
|
8 |
from .pretrainer import GeneformerPretrainer
|
9 |
from .collator_for_gene_classification import DataCollatorForGeneClassification
|
10 |
from .collator_for_cell_classification import DataCollatorForCellClassification
|
11 |
+
from .in_silico_perturber import InSilicoPerturber
|
12 |
+
from .in_silico_perturber_stats import InSilicoPerturberStats
|
geneformer/in_silico_perturber.py
ADDED
@@ -0,0 +1,777 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
"""
|
2 |
+
Geneformer in silico perturber.
|
3 |
+
|
4 |
+
Usage:
|
5 |
+
from geneformer import InSilicoPerturber
|
6 |
+
isp = InSilicoPerturber(perturb_type="delete",
|
7 |
+
perturb_rank_shift=None,
|
8 |
+
genes_to_perturb="all",
|
9 |
+
combos=0,
|
10 |
+
anchor_gene=None,
|
11 |
+
model_type="Pretrained",
|
12 |
+
num_classes=0,
|
13 |
+
emb_mode="cell",
|
14 |
+
cell_emb_style="mean_pool",
|
15 |
+
filter_data={"cell_type":["cardiomyocyte"]},
|
16 |
+
cell_states_to_model={"disease":(["dcm"],["ctrl"],["hcm"])},
|
17 |
+
max_ncells=None,
|
18 |
+
emb_layer=-1,
|
19 |
+
forward_batch_size=100,
|
20 |
+
nproc=4,
|
21 |
+
save_raw_data=False)
|
22 |
+
isp.perturb_data("path/to/model",
|
23 |
+
"path/to/input_data",
|
24 |
+
"path/to/output_directory",
|
25 |
+
"output_prefix")
|
26 |
+
"""
|
27 |
+
|
28 |
+
# imports
|
29 |
+
import itertools as it
|
30 |
+
import logging
|
31 |
+
import pickle
|
32 |
+
import seaborn as sns; sns.set()
|
33 |
+
import torch
|
34 |
+
from collections import defaultdict
|
35 |
+
from datasets import Dataset, load_from_disk
|
36 |
+
from tqdm.notebook import trange
|
37 |
+
from transformers import BertForMaskedLM, BertForTokenClassification, BertForSequenceClassification
|
38 |
+
|
39 |
+
from .tokenizer import TOKEN_DICTIONARY_FILE
|
40 |
+
|
41 |
+
logger = logging.getLogger(__name__)
|
42 |
+
|
43 |
+
def quant_layers(model):
|
44 |
+
layer_nums = []
|
45 |
+
for name, parameter in model.named_parameters():
|
46 |
+
if "layer" in name:
|
47 |
+
layer_nums += [name.split("layer.")[1].split(".")[0]]
|
48 |
+
return int(max(layer_nums))+1
|
49 |
+
|
50 |
+
def flatten_list(megalist):
|
51 |
+
return [item for sublist in megalist for item in sublist]
|
52 |
+
|
53 |
+
def forward_pass_single_cell(model, example_cell, layer_to_quant):
|
54 |
+
example_cell.set_format(type="torch")
|
55 |
+
input_data = example_cell["input_ids"]
|
56 |
+
with torch.no_grad():
|
57 |
+
outputs = model(
|
58 |
+
input_ids = input_data.to("cuda")
|
59 |
+
)
|
60 |
+
emb = torch.squeeze(outputs.hidden_states[layer_to_quant])
|
61 |
+
del outputs
|
62 |
+
return emb
|
63 |
+
|
64 |
+
def perturb_emb_by_index(emb, indices):
|
65 |
+
mask = torch.ones(emb.numel(), dtype=torch.bool)
|
66 |
+
mask[indices] = False
|
67 |
+
return emb[mask]
|
68 |
+
|
69 |
+
def delete_index(example):
|
70 |
+
indexes = example["perturb_index"]
|
71 |
+
if len(indexes)>1:
|
72 |
+
indexes = flatten_list(indexes)
|
73 |
+
for index in sorted(indexes, reverse=True):
|
74 |
+
del example["input_ids"][index]
|
75 |
+
return example
|
76 |
+
|
77 |
+
def overexpress_index(example):
|
78 |
+
indexes = example["perturb_index"]
|
79 |
+
if len(indexes)>1:
|
80 |
+
indexes = flatten_list(indexes)
|
81 |
+
for index in sorted(indexes, reverse=True):
|
82 |
+
example["input_ids"].insert(0, example["input_ids"].pop(index))
|
83 |
+
return example
|
84 |
+
|
85 |
+
def make_perturbation_batch(example_cell,
|
86 |
+
perturb_type,
|
87 |
+
tokens_to_perturb,
|
88 |
+
anchor_token,
|
89 |
+
combo_lvl,
|
90 |
+
num_proc):
|
91 |
+
if tokens_to_perturb == "all":
|
92 |
+
if perturb_type in ["overexpress","activate"]:
|
93 |
+
range_start = 1
|
94 |
+
elif perturb_type in ["delete","inhibit"]:
|
95 |
+
range_start = 0
|
96 |
+
indices_to_perturb = [[i] for i in range(range_start,example_cell["length"][0])]
|
97 |
+
elif combo_lvl>0 and (anchor_token is not None):
|
98 |
+
example_input_ids = example_cell["input_ids "][0]
|
99 |
+
anchor_index = example_input_ids.index(anchor_token[0])
|
100 |
+
indices_to_perturb = [sorted([anchor_index,i]) if i!=anchor_index else None for i in range(example_cell["length"][0])]
|
101 |
+
indices_to_perturb = [item for item in indices_to_perturb if item is not None]
|
102 |
+
else:
|
103 |
+
example_input_ids = example_cell["input_ids"][0]
|
104 |
+
indices_to_perturb = [[example_input_ids.index(token)] if token in example_input_ids else None for token in tokens_to_perturb]
|
105 |
+
indices_to_perturb = [item for item in indices_to_perturb if item is not None]
|
106 |
+
|
107 |
+
# create all permutations of combo_lvl of modifiers from tokens_to_perturb
|
108 |
+
if combo_lvl>0 and (anchor_token is None):
|
109 |
+
if tokens_to_perturb != "all":
|
110 |
+
if len(tokens_to_perturb) == combo_lvl+1:
|
111 |
+
indices_to_perturb = [list(x) for x in it.combinations(indices_to_perturb, combo_lvl+1)]
|
112 |
+
else:
|
113 |
+
all_indices = [[i] for i in range(example_cell["length"][0])]
|
114 |
+
all_indices = [index for index in all_indices if index not in indices_to_perturb]
|
115 |
+
indices_to_perturb = [[[j for i in indices_to_perturb for j in i], x] for x in all_indices]
|
116 |
+
length = len(indices_to_perturb)
|
117 |
+
perturbation_dataset = Dataset.from_dict({"input_ids": example_cell["input_ids"]*length, "perturb_index": indices_to_perturb})
|
118 |
+
if length<400:
|
119 |
+
num_proc_i = 1
|
120 |
+
else:
|
121 |
+
num_proc_i = num_proc
|
122 |
+
if perturb_type == "delete":
|
123 |
+
perturbation_dataset = perturbation_dataset.map(delete_index, num_proc=num_proc_i)
|
124 |
+
elif perturb_type == "overexpress":
|
125 |
+
perturbation_dataset = perturbation_dataset.map(overexpress_index, num_proc=num_proc_i)
|
126 |
+
return perturbation_dataset, indices_to_perturb
|
127 |
+
|
128 |
+
# original cell emb removing the respective perturbed gene emb
|
129 |
+
def make_comparison_batch(original_emb, indices_to_perturb):
|
130 |
+
all_embs_list = []
|
131 |
+
for indices in indices_to_perturb:
|
132 |
+
emb_list = []
|
133 |
+
start = 0
|
134 |
+
if len(indices)>1 and isinstance(indices[0],list):
|
135 |
+
indices = flatten_list(indices)
|
136 |
+
for i in sorted(indices):
|
137 |
+
emb_list += [original_emb[start:i]]
|
138 |
+
start = i+1
|
139 |
+
emb_list += [original_emb[start:]]
|
140 |
+
all_embs_list += [torch.cat(emb_list)]
|
141 |
+
return torch.stack(all_embs_list)
|
142 |
+
|
143 |
+
# average embedding position of goal cell states
|
144 |
+
def get_cell_state_avg_embs(model,
|
145 |
+
filtered_input_data,
|
146 |
+
cell_states_to_model,
|
147 |
+
layer_to_quant,
|
148 |
+
token_dictionary,
|
149 |
+
forward_batch_size,
|
150 |
+
num_proc):
|
151 |
+
possible_states = [value[0]+value[1]+value[2] for value in cell_states_to_model.values()][0]
|
152 |
+
state_embs_dict = dict()
|
153 |
+
for possible_state in possible_states:
|
154 |
+
state_embs_list = []
|
155 |
+
|
156 |
+
def filter_states(example):
|
157 |
+
return example[list(cell_states_to_model.keys())[0]] in [possible_state]
|
158 |
+
filtered_input_data_state = filtered_input_data.filter(filter_states, num_proc=num_proc)
|
159 |
+
total_batch_length = len(filtered_input_data_state)
|
160 |
+
if ((total_batch_length-1)/forward_batch_size).is_integer():
|
161 |
+
forward_batch_size = forward_batch_size-1
|
162 |
+
max_len = max(filtered_input_data_state["length"])
|
163 |
+
for i in range(0, total_batch_length, forward_batch_size):
|
164 |
+
max_range = min(i+forward_batch_size, total_batch_length)
|
165 |
+
|
166 |
+
state_minibatch = filtered_input_data_state.select([i for i in range(i, max_range)])
|
167 |
+
state_minibatch.set_format(type="torch")
|
168 |
+
|
169 |
+
input_data_minibatch = state_minibatch["input_ids"]
|
170 |
+
input_data_minibatch = pad_tensor_list(input_data_minibatch, max_len, token_dictionary)
|
171 |
+
|
172 |
+
with torch.no_grad():
|
173 |
+
outputs = model(
|
174 |
+
input_ids = input_data_minibatch.to("cuda")
|
175 |
+
)
|
176 |
+
|
177 |
+
state_embs_i = outputs.hidden_states[layer_to_quant]
|
178 |
+
state_embs_list += [state_embs_i]
|
179 |
+
del outputs
|
180 |
+
del state_minibatch
|
181 |
+
del input_data_minibatch
|
182 |
+
del state_embs_i
|
183 |
+
torch.cuda.empty_cache()
|
184 |
+
state_embs_stack = torch.cat(state_embs_list)
|
185 |
+
avg_state_emb = torch.mean(state_embs_stack,dim=[0,1],keepdim=True)
|
186 |
+
state_embs_dict[possible_state] = avg_state_emb
|
187 |
+
return state_embs_dict
|
188 |
+
|
189 |
+
# quantify cosine similarity of perturbed vs original or alternate states
|
190 |
+
def quant_cos_sims(model,
|
191 |
+
perturbation_batch,
|
192 |
+
forward_batch_size,
|
193 |
+
layer_to_quant,
|
194 |
+
original_emb,
|
195 |
+
indices_to_perturb,
|
196 |
+
cell_states_to_model,
|
197 |
+
state_embs_dict):
|
198 |
+
cos = torch.nn.CosineSimilarity(dim=2)
|
199 |
+
total_batch_length = len(perturbation_batch)
|
200 |
+
if ((total_batch_length-1)/forward_batch_size).is_integer():
|
201 |
+
forward_batch_size = forward_batch_size-1
|
202 |
+
if cell_states_to_model is None:
|
203 |
+
comparison_batch = make_comparison_batch(original_emb, indices_to_perturb)
|
204 |
+
cos_sims = []
|
205 |
+
else:
|
206 |
+
possible_states = [value[0]+value[1]+value[2] for value in cell_states_to_model.values()][0]
|
207 |
+
cos_sims_vs_alt_dict = dict(zip(possible_states,[[] for i in range(len(possible_states))]))
|
208 |
+
for i in range(0, total_batch_length, forward_batch_size):
|
209 |
+
max_range = min(i+forward_batch_size, total_batch_length)
|
210 |
+
|
211 |
+
perturbation_minibatch = perturbation_batch.select([i for i in range(i, max_range)])
|
212 |
+
perturbation_minibatch.set_format(type="torch")
|
213 |
+
|
214 |
+
input_data_minibatch = perturbation_minibatch["input_ids"]
|
215 |
+
|
216 |
+
with torch.no_grad():
|
217 |
+
outputs = model(
|
218 |
+
input_ids = input_data_minibatch.to("cuda")
|
219 |
+
)
|
220 |
+
del input_data_minibatch
|
221 |
+
del perturbation_minibatch
|
222 |
+
# cosine similarity between original emb and batch items
|
223 |
+
if len(indices_to_perturb)>1:
|
224 |
+
minibatch_emb = torch.squeeze(outputs.hidden_states[layer_to_quant])
|
225 |
+
else:
|
226 |
+
minibatch_emb = outputs.hidden_states[layer_to_quant]
|
227 |
+
if cell_states_to_model is None:
|
228 |
+
minibatch_comparison = comparison_batch[i:max_range]
|
229 |
+
cos_sims += [cos(minibatch_emb, minibatch_comparison).to("cpu")]
|
230 |
+
else:
|
231 |
+
for state in possible_states:
|
232 |
+
cos_sims_vs_alt_dict[state] += cos_sim_shift(original_emb, minibatch_emb, state_embs_dict[state])
|
233 |
+
del outputs
|
234 |
+
del minibatch_emb
|
235 |
+
if cell_states_to_model is None:
|
236 |
+
del minibatch_comparison
|
237 |
+
torch.cuda.empty_cache()
|
238 |
+
if cell_states_to_model is None:
|
239 |
+
cos_sims_stack = torch.cat(cos_sims)
|
240 |
+
return cos_sims_stack
|
241 |
+
else:
|
242 |
+
for state in possible_states:
|
243 |
+
cos_sims_vs_alt_dict[state] = torch.cat(cos_sims_vs_alt_dict[state])
|
244 |
+
return cos_sims_vs_alt_dict
|
245 |
+
|
246 |
+
# calculate cos sim shift of perturbation with respect to origin and alternative cell
|
247 |
+
def cos_sim_shift(original_emb, minibatch_emb, alt_emb):
|
248 |
+
cos = torch.nn.CosineSimilarity(dim=2)
|
249 |
+
original_emb = torch.mean(original_emb,dim=0,keepdim=True)[None, :]
|
250 |
+
alt_emb = alt_emb[None, None, :]
|
251 |
+
origin_v_end = cos(original_emb,alt_emb)
|
252 |
+
perturb_v_end = cos(torch.mean(minibatch_emb,dim=1,keepdim=True),alt_emb)
|
253 |
+
return [(perturb_v_end-origin_v_end).to("cpu")]
|
254 |
+
|
255 |
+
# pad list of tensors and convert to tensor
|
256 |
+
def pad_tensor_list(tensor_list, dynamic_or_constant, token_dictionary):
|
257 |
+
|
258 |
+
pad_token_id = token_dictionary.get("<pad>")
|
259 |
+
|
260 |
+
# Determine maximum tensor length
|
261 |
+
if dynamic_or_constant == "dynamic":
|
262 |
+
max_len = max([tensor.squeeze().numel() for tensor in tensor_list])
|
263 |
+
elif type(dynamic_or_constant) == int:
|
264 |
+
max_len = dynamic_or_constant
|
265 |
+
else:
|
266 |
+
logger.warning(
|
267 |
+
"If padding style is constant, must provide integer value. " \
|
268 |
+
"Setting padding to max input size 2048.")
|
269 |
+
|
270 |
+
# pad all tensors to maximum length
|
271 |
+
tensor_list = [torch.nn.functional.pad(tensor, pad=(0,
|
272 |
+
max_len - tensor.numel()),
|
273 |
+
mode='constant',
|
274 |
+
value=pad_token_id) for tensor in tensor_list]
|
275 |
+
|
276 |
+
# return stacked tensors
|
277 |
+
return torch.stack(tensor_list)
|
278 |
+
|
279 |
+
class InSilicoPerturber:
|
280 |
+
valid_option_dict = {
|
281 |
+
"perturb_type": {"delete","overexpress","inhibit","activate"},
|
282 |
+
"perturb_rank_shift": {None, int},
|
283 |
+
"genes_to_perturb": {"all", list},
|
284 |
+
"combos": {0,1,2},
|
285 |
+
"anchor_gene": {None, str},
|
286 |
+
"model_type": {"Pretrained","GeneClassifier","CellClassifier"},
|
287 |
+
"num_classes": {int},
|
288 |
+
"emb_mode": {"cell","cell_and_gene"},
|
289 |
+
"cell_emb_style": {"mean_pool"},
|
290 |
+
"filter_data": {None, dict},
|
291 |
+
"cell_states_to_model": {None, dict},
|
292 |
+
"max_ncells": {None, int},
|
293 |
+
"emb_layer": {-1, 0},
|
294 |
+
"forward_batch_size": {int},
|
295 |
+
"nproc": {int},
|
296 |
+
"save_raw_data": {False, True},
|
297 |
+
}
|
298 |
+
def __init__(
|
299 |
+
self,
|
300 |
+
perturb_type="delete",
|
301 |
+
perturb_rank_shift=None,
|
302 |
+
genes_to_perturb="all",
|
303 |
+
combos=0,
|
304 |
+
anchor_gene=None,
|
305 |
+
model_type="Pretrained",
|
306 |
+
num_classes=0,
|
307 |
+
emb_mode="cell",
|
308 |
+
cell_emb_style="mean_pool",
|
309 |
+
filter_data=None,
|
310 |
+
cell_states_to_model=None,
|
311 |
+
max_ncells=None,
|
312 |
+
emb_layer=-1,
|
313 |
+
forward_batch_size=100,
|
314 |
+
nproc=4,
|
315 |
+
save_raw_data=False,
|
316 |
+
token_dictionary_file=TOKEN_DICTIONARY_FILE,
|
317 |
+
):
|
318 |
+
"""
|
319 |
+
Initialize in silico perturber.
|
320 |
+
|
321 |
+
Parameters
|
322 |
+
----------
|
323 |
+
perturb_type : {"delete","overexpress","inhibit","activate"}
|
324 |
+
Type of perturbation.
|
325 |
+
"delete": delete gene from rank value encoding
|
326 |
+
"overexpress": move gene to front of rank value encoding
|
327 |
+
"inhibit": move gene to lower quartile of rank value encoding
|
328 |
+
"activate": move gene to higher quartile of rank value encoding
|
329 |
+
perturb_rank_shift : None, int
|
330 |
+
Number of quartiles by which to shift rank of gene.
|
331 |
+
For example, if perturb_type="activate" and perturb_rank_shift=1:
|
332 |
+
genes in 4th quartile will move to middle of 3rd quartile.
|
333 |
+
genes in 3rd quartile will move to middle of 2nd quartile.
|
334 |
+
genes in 2nd quartile will move to middle of 1st quartile.
|
335 |
+
genes in 1st quartile will move to front of rank value encoding.
|
336 |
+
For example, if perturb_type="inhibit" and perturb_rank_shift=2:
|
337 |
+
genes in 1st quartile will move to middle of 3rd quartile.
|
338 |
+
genes in 2nd quartile will move to middle of 4th quartile.
|
339 |
+
genes in 3rd or 4th quartile will move to bottom of rank value encoding.
|
340 |
+
genes_to_perturb : "all", list
|
341 |
+
Default is perturbing each gene detected in each cell in the dataset.
|
342 |
+
Otherwise, may provide a list of ENSEMBL IDs of genes to perturb.
|
343 |
+
combos : {0,1,2}
|
344 |
+
Whether to perturb genes individually (0), in pairs (1), or in triplets (2).
|
345 |
+
anchor_gene : None, str
|
346 |
+
ENSEMBL ID of gene to use as anchor in combination perturbations.
|
347 |
+
For example, if combos=1 and anchor_gene="ENSG00000148400":
|
348 |
+
anchor gene will be perturbed in combination with each other gene.
|
349 |
+
model_type : {"Pretrained","GeneClassifier","CellClassifier"}
|
350 |
+
Whether model is the pretrained Geneformer or a fine-tuned gene or cell classifier.
|
351 |
+
num_classes : int
|
352 |
+
If model is a gene or cell classifier, specify number of classes it was trained to classify.
|
353 |
+
For the pretrained Geneformer model, number of classes is 0 as it is not a classifier.
|
354 |
+
emb_mode : {"cell","cell_and_gene"}
|
355 |
+
Whether to output impact of perturbation on cell and/or gene embeddings.
|
356 |
+
cell_emb_style : "mean_pool"
|
357 |
+
Method for summarizing cell embeddings.
|
358 |
+
Currently only option is mean pooling of gene embeddings for given cell.
|
359 |
+
filter_data : None, dict
|
360 |
+
Default is to use all input data for in silico perturbation study.
|
361 |
+
Otherwise, dictionary specifying .dataset column name and list of values to filter by.
|
362 |
+
cell_states_to_model: None, dict
|
363 |
+
Cell states to model if testing perturbations that achieve goal state change.
|
364 |
+
Single-item dictionary with key being cell attribute (e.g. "disease").
|
365 |
+
Value is tuple of three lists indicating start state, goal end state, and alternate possible end states.
|
366 |
+
max_ncells : None, int
|
367 |
+
Maximum number of cells to test.
|
368 |
+
If None, will test all cells.
|
369 |
+
emb_layer : {-1, 0}
|
370 |
+
Embedding layer to use for quantification.
|
371 |
+
-1: 2nd to last layer (recommended for pretrained Geneformer)
|
372 |
+
0: last layer (recommended for cell classifier fine-tuned for disease state)
|
373 |
+
forward_batch_size : int
|
374 |
+
Batch size for forward pass.
|
375 |
+
nproc : int
|
376 |
+
Number of CPU processes to use.
|
377 |
+
save_raw_data: {False,True}
|
378 |
+
Whether to save raw perturbation data for each gene/cell.
|
379 |
+
token_dictionary_file : Path
|
380 |
+
Path to pickle file containing token dictionary (Ensembl ID:token).
|
381 |
+
"""
|
382 |
+
|
383 |
+
self.perturb_type = perturb_type
|
384 |
+
self.perturb_rank_shift = perturb_rank_shift
|
385 |
+
self.genes_to_perturb = genes_to_perturb
|
386 |
+
self.combos = combos
|
387 |
+
self.anchor_gene = anchor_gene
|
388 |
+
self.model_type = model_type
|
389 |
+
self.num_classes = num_classes
|
390 |
+
self.emb_mode = emb_mode
|
391 |
+
self.cell_emb_style = cell_emb_style
|
392 |
+
self.filter_data = filter_data
|
393 |
+
self.cell_states_to_model = cell_states_to_model
|
394 |
+
self.max_ncells = max_ncells
|
395 |
+
self.emb_layer = emb_layer
|
396 |
+
self.forward_batch_size = forward_batch_size
|
397 |
+
self.nproc = nproc
|
398 |
+
self.save_raw_data = save_raw_data
|
399 |
+
|
400 |
+
self.validate_options()
|
401 |
+
|
402 |
+
# load token dictionary (Ensembl IDs:token)
|
403 |
+
with open(token_dictionary_file, "rb") as f:
|
404 |
+
self.gene_token_dict = pickle.load(f)
|
405 |
+
|
406 |
+
if anchor_gene is None:
|
407 |
+
self.anchor_token = None
|
408 |
+
else:
|
409 |
+
self.anchor_token = self.gene_token_dict[self.anchor_gene]
|
410 |
+
|
411 |
+
if genes_to_perturb == "all":
|
412 |
+
self.tokens_to_perturb = "all"
|
413 |
+
else:
|
414 |
+
self.tokens_to_perturb = [self.gene_token_dict[gene] for gene in self.genes_to_perturb]
|
415 |
+
|
416 |
+
def validate_options(self):
|
417 |
+
for attr_name,valid_options in self.valid_option_dict.items():
|
418 |
+
attr_value = self.__dict__[attr_name]
|
419 |
+
if type(attr_value) not in {list, dict}:
|
420 |
+
if attr_value in valid_options:
|
421 |
+
continue
|
422 |
+
valid_type = False
|
423 |
+
for option in valid_options:
|
424 |
+
if (option in [int,list,dict]) and isinstance(attr_value, option):
|
425 |
+
valid_type = True
|
426 |
+
break
|
427 |
+
if valid_type:
|
428 |
+
continue
|
429 |
+
logger.error(
|
430 |
+
f"Invalid option for {attr_name}. " \
|
431 |
+
f"Valid options for {attr_name}: {valid_options}"
|
432 |
+
)
|
433 |
+
raise
|
434 |
+
|
435 |
+
if self.perturb_type in ["delete","overexpress"]:
|
436 |
+
if self.perturb_rank_shift is not None:
|
437 |
+
if self.perturb_type == "delete":
|
438 |
+
logger.warning(
|
439 |
+
"perturb_rank_shift set to None. " \
|
440 |
+
"If perturb type is delete then gene is deleted entirely " \
|
441 |
+
"rather than shifted by quartile")
|
442 |
+
elif self.perturb_type == "overexpress":
|
443 |
+
logger.warning(
|
444 |
+
"perturb_rank_shift set to None. " \
|
445 |
+
"If perturb type is activate then gene is moved to front " \
|
446 |
+
"of rank value encoding rather than shifted by quartile")
|
447 |
+
self.perturb_rank_shift = None
|
448 |
+
|
449 |
+
if (self.anchor_gene is not None) and (self.emb_mode == "cell_and_gene"):
|
450 |
+
self.emb_mode = "cell"
|
451 |
+
logger.warning(
|
452 |
+
"emb_mode set to 'cell'. " \
|
453 |
+
"Currently, analysis with anchor gene " \
|
454 |
+
"only outputs effect on cell embeddings.")
|
455 |
+
|
456 |
+
if self.cell_states_to_model is not None:
|
457 |
+
if (len(self.cell_states_to_model.items()) == 1):
|
458 |
+
for key,value in self.cell_states_to_model.items():
|
459 |
+
if (len(value) == 3) and isinstance(value, tuple):
|
460 |
+
if isinstance(value[0],list) and isinstance(value[1],list) and isinstance(value[2],list):
|
461 |
+
if len(value[0]) == 1 and len(value[1]) == 1:
|
462 |
+
all_values = value[0]+value[1]+value[2]
|
463 |
+
if len(all_values) == len(set(all_values)):
|
464 |
+
continue
|
465 |
+
else:
|
466 |
+
logger.error(
|
467 |
+
"Cell states to model must be a single-item dictionary with " \
|
468 |
+
"key being cell attribute (e.g. 'disease') and value being " \
|
469 |
+
"tuple of three lists indicating start state, goal end state, and alternate possible end states. " \
|
470 |
+
"Values should all be unique. " \
|
471 |
+
"For example: {'disease':(['dcm'],['ctrl'],['hcm'])}")
|
472 |
+
raise
|
473 |
+
if self.anchor_gene is not None:
|
474 |
+
self.anchor_gene = None
|
475 |
+
logger.warning(
|
476 |
+
"anchor_gene set to None. " \
|
477 |
+
"Currently, anchor gene not available " \
|
478 |
+
"when modeling multiple cell states.")
|
479 |
+
|
480 |
+
if self.perturb_type in ["inhibit","activate"]:
|
481 |
+
if self.perturb_rank_shift is None:
|
482 |
+
logger.error(
|
483 |
+
"If perturb type is inhibit or activate then " \
|
484 |
+
"quartile to shift by must be specified.")
|
485 |
+
raise
|
486 |
+
|
487 |
+
for key,value in self.filter_data.items():
|
488 |
+
if type(value) != list:
|
489 |
+
self.filter_data[key] = [value]
|
490 |
+
logger.warning(
|
491 |
+
"Values in filter_data dict must be lists. " \
|
492 |
+
f"Changing {key} value to list ([{value}]).")
|
493 |
+
|
494 |
+
def perturb_data(self,
|
495 |
+
model_directory,
|
496 |
+
input_data_file,
|
497 |
+
output_directory,
|
498 |
+
output_prefix):
|
499 |
+
"""
|
500 |
+
Perturb genes in input data and save as results in output_directory.
|
501 |
+
|
502 |
+
Parameters
|
503 |
+
----------
|
504 |
+
model_directory : Path
|
505 |
+
Path to directory containing model
|
506 |
+
input_data_file : Path
|
507 |
+
Path to directory containing .dataset inputs
|
508 |
+
output_directory : Path
|
509 |
+
Path to directory where perturbation data will be saved as .csv
|
510 |
+
output_prefix : str
|
511 |
+
Prefix for output .dataset
|
512 |
+
"""
|
513 |
+
|
514 |
+
filtered_input_data = self.load_and_filter(input_data_file)
|
515 |
+
model = self.load_model(model_directory)
|
516 |
+
layer_to_quant = quant_layers(model)+self.emb_layer
|
517 |
+
|
518 |
+
if self.cell_states_to_model is None:
|
519 |
+
state_embs_dict = None
|
520 |
+
else:
|
521 |
+
# get dictionary of average cell state embeddings for comparison
|
522 |
+
state_embs_dict = get_cell_state_avg_embs(model,
|
523 |
+
filtered_input_data,
|
524 |
+
self.cell_states_to_model,
|
525 |
+
layer_to_quant,
|
526 |
+
self.gene_token_dict,
|
527 |
+
self.forward_batch_size,
|
528 |
+
self.nproc)
|
529 |
+
self.in_silico_perturb(model,
|
530 |
+
filtered_input_data,
|
531 |
+
layer_to_quant,
|
532 |
+
state_embs_dict,
|
533 |
+
output_directory,
|
534 |
+
output_prefix)
|
535 |
+
|
536 |
+
# if self.save_raw_data is False:
|
537 |
+
# # delete intermediate dictionaries
|
538 |
+
# output_dir = os.listdir(output_directory)
|
539 |
+
# for output_file in output_dir:
|
540 |
+
# if output_file.endswith("_raw.pickle"):
|
541 |
+
# os.remove(os.path.join(output_directory, output_file))
|
542 |
+
|
543 |
+
# load data and filter by defined criteria
|
544 |
+
def load_and_filter(self, input_data_file):
|
545 |
+
data = load_from_disk(input_data_file)
|
546 |
+
for key,value in self.filter_data.items():
|
547 |
+
def filter_data(example):
|
548 |
+
return example[key] in value
|
549 |
+
data = data.filter(filter_data, num_proc=self.nproc)
|
550 |
+
if len(data) == 0:
|
551 |
+
logger.error(
|
552 |
+
"No cells remain after filtering. Check filtering criteria.")
|
553 |
+
raise
|
554 |
+
data_shuffled = data.shuffle(seed=42)
|
555 |
+
num_cells = len(data_shuffled)
|
556 |
+
# if max number of cells is defined, then subsample to this max number
|
557 |
+
if self.max_ncells != None:
|
558 |
+
num_cells = min(self.max_ncells,num_cells)
|
559 |
+
data_subset = data_shuffled.select([i for i in range(num_cells)])
|
560 |
+
# sort dataset with largest cell first to encounter any memory errors earlier
|
561 |
+
data_sorted = data_subset.sort("length",reverse=True)
|
562 |
+
return data_sorted
|
563 |
+
|
564 |
+
# load model to GPU
|
565 |
+
def load_model(self, model_directory):
|
566 |
+
if self.model_type == "Pretrained":
|
567 |
+
model = BertForMaskedLM.from_pretrained(model_directory,
|
568 |
+
output_hidden_states=True,
|
569 |
+
output_attentions=False)
|
570 |
+
elif self.model_type == "GeneClassifier":
|
571 |
+
model = BertForTokenClassification.from_pretrained(model_directory,
|
572 |
+
num_labels=self.num_classes,
|
573 |
+
output_hidden_states=True,
|
574 |
+
output_attentions=False)
|
575 |
+
elif self.model_type == "CellClassifier":
|
576 |
+
model = BertForSequenceClassification.from_pretrained(model_directory,
|
577 |
+
num_labels=self.num_classes,
|
578 |
+
output_hidden_states=True,
|
579 |
+
output_attentions=False)
|
580 |
+
# put the model in eval mode for fwd pass
|
581 |
+
model.eval()
|
582 |
+
model = model.to("cuda:0")
|
583 |
+
return model
|
584 |
+
|
585 |
+
# determine effect of perturbation on other genes
|
586 |
+
def in_silico_perturb(self,
|
587 |
+
model,
|
588 |
+
filtered_input_data,
|
589 |
+
layer_to_quant,
|
590 |
+
state_embs_dict,
|
591 |
+
output_directory,
|
592 |
+
output_prefix):
|
593 |
+
|
594 |
+
output_path_prefix = f"{output_directory}in_silico_{self.perturb_type}_{output_prefix}_dict_1Kbatch"
|
595 |
+
|
596 |
+
# filter dataset for cells that have tokens to be perturbed
|
597 |
+
if self.anchor_token is not None:
|
598 |
+
def if_has_tokens_to_perturb(example):
|
599 |
+
return (len(set(example["input_ids"]).intersection(self.anchor_token))==len(self.anchor_token))
|
600 |
+
filtered_input_data = filtered_input_data.filter(if_has_tokens_to_perturb, num_proc=self.nproc)
|
601 |
+
logger.info(f"# cells with anchor gene: {len(filtered_input_data)}")
|
602 |
+
if self.tokens_to_perturb != "all":
|
603 |
+
def if_has_tokens_to_perturb(example):
|
604 |
+
return (len(set(example["input_ids"]).intersection(self.tokens_to_perturb))>self.combos)
|
605 |
+
filtered_input_data = filtered_input_data.filter(if_has_tokens_to_perturb, num_proc=self.nproc)
|
606 |
+
|
607 |
+
cos_sims_dict = defaultdict(list)
|
608 |
+
pickle_batch = -1
|
609 |
+
|
610 |
+
for i in trange(len(filtered_input_data)):
|
611 |
+
example_cell = filtered_input_data.select([i])
|
612 |
+
original_emb = forward_pass_single_cell(model, example_cell, layer_to_quant)
|
613 |
+
gene_list = torch.squeeze(example_cell["input_ids"])
|
614 |
+
|
615 |
+
# reset to original type to prevent downstream issues due to forward_pass_single_cell modifying as torch format in place
|
616 |
+
example_cell = filtered_input_data.select([i])
|
617 |
+
|
618 |
+
if self.anchor_token is None:
|
619 |
+
for combo_lvl in range(self.combos+1):
|
620 |
+
perturbation_batch, indices_to_perturb = make_perturbation_batch(example_cell,
|
621 |
+
self.perturb_type,
|
622 |
+
self.tokens_to_perturb,
|
623 |
+
self.anchor_token,
|
624 |
+
combo_lvl,
|
625 |
+
self.nproc)
|
626 |
+
cos_sims_data = quant_cos_sims(model,
|
627 |
+
perturbation_batch,
|
628 |
+
self.forward_batch_size,
|
629 |
+
layer_to_quant,
|
630 |
+
original_emb,
|
631 |
+
indices_to_perturb,
|
632 |
+
self.cell_states_to_model,
|
633 |
+
state_embs_dict)
|
634 |
+
|
635 |
+
if self.cell_states_to_model is None:
|
636 |
+
# update cos sims dict
|
637 |
+
# key is tuple of (perturbed_gene, affected_gene)
|
638 |
+
# or (perturbed_gene, "cell_emb") for avg cell emb change
|
639 |
+
cos_sims_data = cos_sims_data.to("cuda")
|
640 |
+
for j in range(cos_sims_data.shape[0]):
|
641 |
+
if self.genes_to_perturb != "all":
|
642 |
+
j_index = torch.tensor(indices_to_perturb[j])
|
643 |
+
if j_index.shape[0]>1:
|
644 |
+
j_index = torch.squeeze(j_index)
|
645 |
+
else:
|
646 |
+
j_index = torch.tensor([j])
|
647 |
+
perturbed_gene = torch.index_select(gene_list, 0, j_index)
|
648 |
+
|
649 |
+
if perturbed_gene.shape[0]==1:
|
650 |
+
perturbed_gene = perturbed_gene.item()
|
651 |
+
elif perturbed_gene.shape[0]>1:
|
652 |
+
perturbed_gene = tuple(perturbed_gene.tolist())
|
653 |
+
|
654 |
+
cell_cos_sim = torch.mean(cos_sims_data[j]).item()
|
655 |
+
cos_sims_dict[(perturbed_gene, "cell_emb")] += [cell_cos_sim]
|
656 |
+
|
657 |
+
# not_j_index = list(set(i for i in range(gene_list.shape[0])).difference(j_index))
|
658 |
+
# gene_list_j = torch.index_select(gene_list, 0, j_index)
|
659 |
+
if self.emb_mode == "cell_and_gene":
|
660 |
+
for k in range(cos_sims_data.shape[1]):
|
661 |
+
cos_sim_value = cos_sims_data[j][k]
|
662 |
+
affected_gene = gene_list[k].item()
|
663 |
+
cos_sims_dict[(perturbed_gene, affected_gene)] += [cos_sim_value.item()]
|
664 |
+
else:
|
665 |
+
# update cos sims dict
|
666 |
+
# key is tuple of (perturbed_gene, "cell_emb")
|
667 |
+
# value is list of tuples of cos sims for cell_states_to_model
|
668 |
+
origin_state_key = [value[0] for value in self.cell_states_to_model.values()][0][0]
|
669 |
+
cos_sims_origin = cos_sims_data[origin_state_key]
|
670 |
+
|
671 |
+
for j in range(cos_sims_origin.shape[0]):
|
672 |
+
if (self.genes_to_perturb != "all") or (combo_lvl>0):
|
673 |
+
j_index = torch.tensor(indices_to_perturb[j])
|
674 |
+
if j_index.shape[0]>1:
|
675 |
+
j_index = torch.squeeze(j_index)
|
676 |
+
else:
|
677 |
+
j_index = torch.tensor([j])
|
678 |
+
perturbed_gene = torch.index_select(gene_list, 0, j_index)
|
679 |
+
|
680 |
+
if perturbed_gene.shape[0]==1:
|
681 |
+
perturbed_gene = perturbed_gene.item()
|
682 |
+
elif perturbed_gene.shape[0]>1:
|
683 |
+
perturbed_gene = tuple(perturbed_gene.tolist())
|
684 |
+
|
685 |
+
data_list = []
|
686 |
+
for data in list(cos_sims_data.values()):
|
687 |
+
data_item = data.to("cuda")
|
688 |
+
cell_data = torch.mean(data_item[j]).item()
|
689 |
+
data_list += [cell_data]
|
690 |
+
cos_sims_dict[(perturbed_gene, "cell_emb")] += [tuple(data_list)]
|
691 |
+
|
692 |
+
elif self.anchor_token is not None:
|
693 |
+
perturbation_batch, indices_to_perturb = make_perturbation_batch(example_cell,
|
694 |
+
self.perturb_type,
|
695 |
+
self.tokens_to_perturb,
|
696 |
+
None, # first run without anchor token to test individual gene perturbations
|
697 |
+
0,
|
698 |
+
self.nproc)
|
699 |
+
cos_sims_data = quant_cos_sims(model,
|
700 |
+
perturbation_batch,
|
701 |
+
self.forward_batch_size,
|
702 |
+
layer_to_quant,
|
703 |
+
original_emb,
|
704 |
+
indices_to_perturb,
|
705 |
+
self.cell_states_to_model,
|
706 |
+
state_embs_dict)
|
707 |
+
cos_sims_data = cos_sims_data.to("cuda")
|
708 |
+
|
709 |
+
combo_perturbation_batch, combo_indices_to_perturb = make_perturbation_batch(example_cell,
|
710 |
+
self.perturb_type,
|
711 |
+
self.tokens_to_perturb,
|
712 |
+
self.anchor_token,
|
713 |
+
1,
|
714 |
+
self.nproc)
|
715 |
+
combo_cos_sims_data = quant_cos_sims(model,
|
716 |
+
combo_perturbation_batch,
|
717 |
+
self.forward_batch_size,
|
718 |
+
layer_to_quant,
|
719 |
+
original_emb,
|
720 |
+
combo_indices_to_perturb,
|
721 |
+
self.cell_states_to_model,
|
722 |
+
state_embs_dict)
|
723 |
+
combo_cos_sims_data = combo_cos_sims_data.to("cuda")
|
724 |
+
|
725 |
+
# update cos sims dict
|
726 |
+
# key is tuple of (perturbed_gene, "cell_emb") for avg cell emb change
|
727 |
+
anchor_index = example_cell["input_ids"][0].index(self.anchor_token[0])
|
728 |
+
anchor_cell_cos_sim = torch.mean(cos_sims_data[anchor_index]).item()
|
729 |
+
non_anchor_indices = [k for k in range(cos_sims_data.shape[0]) if k != anchor_index]
|
730 |
+
cos_sims_data = cos_sims_data[non_anchor_indices,:]
|
731 |
+
|
732 |
+
for j in range(cos_sims_data.shape[0]):
|
733 |
+
|
734 |
+
if j<anchor_index:
|
735 |
+
j_index = torch.tensor([j])
|
736 |
+
else:
|
737 |
+
j_index = torch.tensor([j+1])
|
738 |
+
|
739 |
+
perturbed_gene = torch.index_select(gene_list, 0, j_index)
|
740 |
+
perturbed_gene = perturbed_gene.item()
|
741 |
+
|
742 |
+
cell_cos_sim = torch.mean(cos_sims_data[j]).item()
|
743 |
+
combo_cos_sim = torch.mean(combo_cos_sims_data[j]).item()
|
744 |
+
cos_sims_dict[(perturbed_gene, "cell_emb")] += [(anchor_cell_cos_sim, # cos sim anchor gene alone
|
745 |
+
cell_cos_sim, # cos sim deleted gene alone
|
746 |
+
combo_cos_sim)] # cos sim anchor gene + deleted gene
|
747 |
+
|
748 |
+
# save dict to disk every 100 cells
|
749 |
+
if (i/100).is_integer():
|
750 |
+
with open(f"{output_path_prefix}{pickle_batch}_raw.pickle", "wb") as fp:
|
751 |
+
pickle.dump(cos_sims_dict, fp)
|
752 |
+
# reset and clear memory every 1000 cells
|
753 |
+
if (i/1000).is_integer():
|
754 |
+
pickle_batch = pickle_batch+1
|
755 |
+
# clear memory
|
756 |
+
del perturbed_gene
|
757 |
+
del cos_sims_data
|
758 |
+
if self.cell_states_to_model is None:
|
759 |
+
del cell_cos_sim
|
760 |
+
if self.cell_states_to_model is not None:
|
761 |
+
del cell_data
|
762 |
+
del data_list
|
763 |
+
elif self.anchor_token is None:
|
764 |
+
del affected_gene
|
765 |
+
del cos_sim_value
|
766 |
+
else:
|
767 |
+
del combo_cos_sim
|
768 |
+
del combo_cos_sims_data
|
769 |
+
# reset dict
|
770 |
+
del cos_sims_dict
|
771 |
+
cos_sims_dict = defaultdict(list)
|
772 |
+
torch.cuda.empty_cache()
|
773 |
+
|
774 |
+
# save remainder cells
|
775 |
+
with open(f"{output_path_prefix}{pickle_batch}_raw.pickle", "wb") as fp:
|
776 |
+
pickle.dump(cos_sims_dict, fp)
|
777 |
+
|
geneformer/in_silico_perturber_stats.py
ADDED
@@ -0,0 +1,302 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
"""
|
2 |
+
Geneformer in silico perturber stats generator.
|
3 |
+
|
4 |
+
Usage:
|
5 |
+
from geneformer import InSilicoPerturberStats
|
6 |
+
ispstats = InSilicoPerturberStats(mode="goal_state_shift",
|
7 |
+
combos=0,
|
8 |
+
anchor_gene=None,
|
9 |
+
cell_states_to_model={"disease":(["dcm"],["ctrl"],["hcm"])})
|
10 |
+
ispstats.get_stats("path/to/input_data",
|
11 |
+
None,
|
12 |
+
"path/to/output_directory",
|
13 |
+
"output_prefix")
|
14 |
+
"""
|
15 |
+
|
16 |
+
|
17 |
+
import os
|
18 |
+
import logging
|
19 |
+
import numpy as np
|
20 |
+
import pandas as pd
|
21 |
+
import pickle
|
22 |
+
import statsmodels.stats.multitest as smt
|
23 |
+
from pathlib import Path
|
24 |
+
from scipy.stats import ranksums
|
25 |
+
from tqdm.notebook import trange
|
26 |
+
|
27 |
+
from .tokenizer import TOKEN_DICTIONARY_FILE
|
28 |
+
|
29 |
+
GENE_NAME_ID_DICTIONARY_FILE = Path(__file__).parent / "gene_name_id_dict.pkl"
|
30 |
+
|
31 |
+
logger = logging.getLogger(__name__)
|
32 |
+
|
33 |
+
# invert dictionary keys/values
|
34 |
+
def invert_dict(dictionary):
|
35 |
+
return {v: k for k, v in dictionary.items()}
|
36 |
+
|
37 |
+
# read raw dictionary files
|
38 |
+
def read_dictionaries(dir, cell_or_gene_emb):
|
39 |
+
dict_list = []
|
40 |
+
for file in os.listdir(dir):
|
41 |
+
# process only _raw.pickle files
|
42 |
+
if file.endswith("_raw.pickle"):
|
43 |
+
with open(f"{dir}/{file}", "rb") as fp:
|
44 |
+
cos_sims_dict = pickle.load(fp)
|
45 |
+
if cell_or_gene_emb == "cell":
|
46 |
+
cell_emb_dict = {k: v for k,
|
47 |
+
v in cos_sims_dict.items() if v and "cell_emb" in k}
|
48 |
+
dict_list += [cell_emb_dict]
|
49 |
+
return dict_list
|
50 |
+
|
51 |
+
# get complete gene list
|
52 |
+
def get_gene_list(dict_list):
|
53 |
+
gene_set = set()
|
54 |
+
for dict_i in dict_list:
|
55 |
+
gene_set.update([k[0] for k, v in dict_i.items() if v])
|
56 |
+
gene_list = list(gene_set)
|
57 |
+
gene_list.sort()
|
58 |
+
return gene_list
|
59 |
+
|
60 |
+
def n_detections(token, dict_list):
|
61 |
+
cos_sim_megalist = []
|
62 |
+
for dict_i in dict_list:
|
63 |
+
cos_sim_megalist += dict_i.get((token, "cell_emb"),[])
|
64 |
+
return len(cos_sim_megalist)
|
65 |
+
|
66 |
+
def get_fdr(pvalues):
|
67 |
+
return list(smt.multipletests(pvalues, alpha=0.05, method="fdr_bh")[1])
|
68 |
+
|
69 |
+
def isp_stats(cos_sims_df, dict_list, cell_states_to_model):
|
70 |
+
|
71 |
+
random_tuples = []
|
72 |
+
for i in trange(cos_sims_df.shape[0]):
|
73 |
+
token = cos_sims_df["Gene"][i]
|
74 |
+
for dict_i in dict_list:
|
75 |
+
random_tuples += dict_i.get((token, "cell_emb"),[])
|
76 |
+
goal_end_random_megalist = [goal_end for goal_end,alt_end,start_state in random_tuples]
|
77 |
+
alt_end_random_megalist = [alt_end for goal_end,alt_end,start_state in random_tuples]
|
78 |
+
start_state_random_megalist = [start_state for goal_end,alt_end,start_state in random_tuples]
|
79 |
+
|
80 |
+
# downsample to improve speed of ranksums
|
81 |
+
if len(goal_end_random_megalist) > 100_000:
|
82 |
+
random.seed(42)
|
83 |
+
goal_end_random_megalist = random.sample(goal_end_random_megalist, k=100_000)
|
84 |
+
if len(alt_end_random_megalist) > 100_000:
|
85 |
+
random.seed(42)
|
86 |
+
alt_end_random_megalist = random.sample(alt_end_random_megalist, k=100_000)
|
87 |
+
if len(start_state_random_megalist) > 100_000:
|
88 |
+
random.seed(42)
|
89 |
+
start_state_random_megalist = random.sample(start_state_random_megalist, k=100_000)
|
90 |
+
|
91 |
+
names=["Gene",
|
92 |
+
"Gene_name",
|
93 |
+
"Ensembl_ID",
|
94 |
+
"Shift_from_goal_end",
|
95 |
+
"Shift_from_alt_end",
|
96 |
+
"Goal_end_vs_random_pval",
|
97 |
+
"Alt_end_vs_random_pval"]
|
98 |
+
cos_sims_full_df = pd.DataFrame(columns=names)
|
99 |
+
|
100 |
+
for i in trange(cos_sims_df.shape[0]):
|
101 |
+
token = cos_sims_df["Gene"][i]
|
102 |
+
name = cos_sims_df["Gene_name"][i]
|
103 |
+
ensembl_id = cos_sims_df["Ensembl_ID"][i]
|
104 |
+
token_tuples = []
|
105 |
+
|
106 |
+
for dict_i in dict_list:
|
107 |
+
token_tuples += dict_i.get((token, "cell_emb"),[])
|
108 |
+
|
109 |
+
goal_end_cos_sim_megalist = [goal_end for goal_end,alt_end,start_state in token_tuples]
|
110 |
+
alt_end_cos_sim_megalist = [alt_end for goal_end,alt_end,start_state in token_tuples]
|
111 |
+
|
112 |
+
mean_goal_end = np.mean(goal_end_cos_sim_megalist)
|
113 |
+
mean_alt_end = np.mean(alt_end_cos_sim_megalist)
|
114 |
+
|
115 |
+
pval_goal_end = ranksums(goal_end_random_megalist,goal_end_cos_sim_megalist).pvalue
|
116 |
+
pval_alt_end = ranksums(alt_end_random_megalist,alt_end_cos_sim_megalist).pvalue
|
117 |
+
|
118 |
+
data_i = [token,
|
119 |
+
name,
|
120 |
+
ensembl_id,
|
121 |
+
mean_goal_end,
|
122 |
+
mean_alt_end,
|
123 |
+
pval_goal_end,
|
124 |
+
pval_alt_end]
|
125 |
+
|
126 |
+
cos_sims_df_i = pd.DataFrame(dict(zip(names,data_i)),index=[i])
|
127 |
+
cos_sims_full_df = pd.concat([cos_sims_full_df,cos_sims_df_i])
|
128 |
+
|
129 |
+
cos_sims_full_df["Goal_end_FDR"] = get_fdr(list(cos_sims_full_df["Goal_end_vs_random_pval"]))
|
130 |
+
cos_sims_full_df["Alt_end_FDR"] = get_fdr(list(cos_sims_full_df["Alt_end_vs_random_pval"]))
|
131 |
+
|
132 |
+
return cos_sims_full_df
|
133 |
+
|
134 |
+
class InSilicoPerturberStats:
|
135 |
+
valid_option_dict = {
|
136 |
+
"mode": {"goal_state_shift","vs_null","vs_random"},
|
137 |
+
"combos": {0,1,2},
|
138 |
+
"anchor_gene": {None, str},
|
139 |
+
"cell_states_to_model": {None, dict},
|
140 |
+
}
|
141 |
+
def __init__(
|
142 |
+
self,
|
143 |
+
mode="vs_random",
|
144 |
+
combos=0,
|
145 |
+
anchor_gene=None,
|
146 |
+
cell_states_to_model=None,
|
147 |
+
token_dictionary_file=TOKEN_DICTIONARY_FILE,
|
148 |
+
gene_name_id_dictionary_file=GENE_NAME_ID_DICTIONARY_FILE,
|
149 |
+
):
|
150 |
+
"""
|
151 |
+
Initialize in silico perturber stats generator.
|
152 |
+
|
153 |
+
Parameters
|
154 |
+
----------
|
155 |
+
mode : {"goal_state_shift","vs_null","vs_random"}
|
156 |
+
Type of stats.
|
157 |
+
"goal_state_shift": perturbation vs. random for desired cell state shift
|
158 |
+
"vs_null": perturbation vs. null from provided null distribution dataset
|
159 |
+
"vs_random": perturbation vs. random gene perturbations in that cell (no goal direction)
|
160 |
+
combos : {0,1,2}
|
161 |
+
Whether to perturb genes individually (0), in pairs (1), or in triplets (2).
|
162 |
+
anchor_gene : None, str
|
163 |
+
ENSEMBL ID of gene to use as anchor in combination perturbations.
|
164 |
+
For example, if combos=1 and anchor_gene="ENSG00000148400":
|
165 |
+
anchor gene will be perturbed in combination with each other gene.
|
166 |
+
cell_states_to_model: None, dict
|
167 |
+
Cell states to model if testing perturbations that achieve goal state change.
|
168 |
+
Single-item dictionary with key being cell attribute (e.g. "disease").
|
169 |
+
Value is tuple of three lists indicating start state, goal end state, and alternate possible end states.
|
170 |
+
token_dictionary_file : Path
|
171 |
+
Path to pickle file containing token dictionary (Ensembl ID:token).
|
172 |
+
gene_name_id_dictionary_file : Path
|
173 |
+
Path to pickle file containing gene name to ID dictionary (gene name:Ensembl ID).
|
174 |
+
"""
|
175 |
+
|
176 |
+
self.mode = mode
|
177 |
+
self.combos = combos
|
178 |
+
self.anchor_gene = anchor_gene
|
179 |
+
self.cell_states_to_model = cell_states_to_model
|
180 |
+
|
181 |
+
self.validate_options()
|
182 |
+
|
183 |
+
# load token dictionary (Ensembl IDs:token)
|
184 |
+
with open(token_dictionary_file, "rb") as f:
|
185 |
+
self.gene_token_dict = pickle.load(f)
|
186 |
+
|
187 |
+
# load gene name dictionary (gene name:Ensembl ID)
|
188 |
+
with open(gene_name_id_dictionary_file, "rb") as f:
|
189 |
+
self.gene_name_id_dict = pickle.load(f)
|
190 |
+
|
191 |
+
if anchor_gene is None:
|
192 |
+
self.anchor_token = None
|
193 |
+
else:
|
194 |
+
self.anchor_token = self.gene_token_dict[self.anchor_gene]
|
195 |
+
|
196 |
+
def validate_options(self):
|
197 |
+
for attr_name,valid_options in self.valid_option_dict.items():
|
198 |
+
attr_value = self.__dict__[attr_name]
|
199 |
+
if type(attr_value) not in {list, dict}:
|
200 |
+
if attr_value in valid_options:
|
201 |
+
continue
|
202 |
+
valid_type = False
|
203 |
+
for option in valid_options:
|
204 |
+
if (option in [int,list,dict]) and isinstance(attr_value, option):
|
205 |
+
valid_type = True
|
206 |
+
break
|
207 |
+
if valid_type:
|
208 |
+
continue
|
209 |
+
logger.error(
|
210 |
+
f"Invalid option for {attr_name}. " \
|
211 |
+
f"Valid options for {attr_name}: {valid_options}"
|
212 |
+
)
|
213 |
+
raise
|
214 |
+
|
215 |
+
if self.cell_states_to_model is not None:
|
216 |
+
if (len(self.cell_states_to_model.items()) == 1):
|
217 |
+
for key,value in self.cell_states_to_model.items():
|
218 |
+
if (len(value) == 3) and isinstance(value, tuple):
|
219 |
+
if isinstance(value[0],list) and isinstance(value[1],list) and isinstance(value[2],list):
|
220 |
+
if len(value[0]) == 1 and len(value[1]) == 1:
|
221 |
+
all_values = value[0]+value[1]+value[2]
|
222 |
+
if len(all_values) == len(set(all_values)):
|
223 |
+
continue
|
224 |
+
else:
|
225 |
+
logger.error(
|
226 |
+
"Cell states to model must be a single-item dictionary with " \
|
227 |
+
"key being cell attribute (e.g. 'disease') and value being " \
|
228 |
+
"tuple of three lists indicating start state, goal end state, and alternate possible end states. " \
|
229 |
+
"Values should all be unique. " \
|
230 |
+
"For example: {'disease':(['start_state'],['ctrl'],['alt_end'])}")
|
231 |
+
raise
|
232 |
+
if self.anchor_gene is not None:
|
233 |
+
self.anchor_gene = None
|
234 |
+
logger.warning(
|
235 |
+
"anchor_gene set to None. " \
|
236 |
+
"Currently, anchor gene not available " \
|
237 |
+
"when modeling multiple cell states.")
|
238 |
+
|
239 |
+
def get_stats(self,
|
240 |
+
input_data_directory,
|
241 |
+
null_dist_data_directory,
|
242 |
+
output_directory,
|
243 |
+
output_prefix):
|
244 |
+
"""
|
245 |
+
Get stats for in silico perturbation data and save as results in output_directory.
|
246 |
+
|
247 |
+
Parameters
|
248 |
+
----------
|
249 |
+
input_data_directory : Path
|
250 |
+
Path to directory containing cos_sim dictionary inputs
|
251 |
+
null_dist_data_directory : Path
|
252 |
+
Path to directory containing null distribution cos_sim dictionary inputs
|
253 |
+
output_directory : Path
|
254 |
+
Path to directory where perturbation data will be saved as .csv
|
255 |
+
output_prefix : str
|
256 |
+
Prefix for output .dataset
|
257 |
+
"""
|
258 |
+
|
259 |
+
self.gene_token_id_dict = invert_dict(self.gene_token_dict)
|
260 |
+
self.gene_id_name_dict = invert_dict(self.gene_name_id_dict)
|
261 |
+
|
262 |
+
if self.mode == "goal_state_shift":
|
263 |
+
dict_list = read_dictionaries(input_data_directory,"cell")
|
264 |
+
else:
|
265 |
+
logger.error(
|
266 |
+
"Currently, only mode available is stats for goal_state_shift.")
|
267 |
+
raise
|
268 |
+
|
269 |
+
# obtain total gene list
|
270 |
+
gene_list = get_gene_list(dict_list)
|
271 |
+
|
272 |
+
# initiate results dataframe
|
273 |
+
cos_sims_df_initial = pd.DataFrame({"Gene": gene_list,
|
274 |
+
"Gene_name": [self.token_to_gene_name(item) \
|
275 |
+
for item in gene_list], \
|
276 |
+
"Ensembl_ID": [self.gene_token_id_dict[genes[1]] \
|
277 |
+
if isinstance(genes,tuple) else \
|
278 |
+
self.gene_token_id_dict[genes] \
|
279 |
+
for genes in gene_list]}, \
|
280 |
+
index=[i for i in range(len(gene_list))])
|
281 |
+
|
282 |
+
# # add ENSEMBL ID for genes
|
283 |
+
# cos_sims_df_initial["Ensembl_ID"] = [self.gene_token_id_dict[genes[1]] if isinstance(genes,tuple) else self.gene_token_id_dict[genes] for genes in list(cos_sims_df_initial["Gene"])]
|
284 |
+
|
285 |
+
cos_sims_df = isp_stats(cos_sims_df_initial, dict_list, self.cell_states_to_model)
|
286 |
+
|
287 |
+
# quantify number of detections of each gene
|
288 |
+
cos_sims_df["N_Detections"] = [n_detections(i, dict_list) for i in cos_sims_df["Gene"]]
|
289 |
+
|
290 |
+
# sort by shift to desired state
|
291 |
+
cos_sims_df = cos_sims_df.sort_values(by=["Shift_from_goal_end",
|
292 |
+
"Goal_end_FDR"])
|
293 |
+
|
294 |
+
# save perturbation stats to output_path
|
295 |
+
output_path = (Path(output_directory) / output_prefix).with_suffix(".csv")
|
296 |
+
cos_sims_df.to_csv(output_path)
|
297 |
+
|
298 |
+
def token_to_gene_name(self, item):
|
299 |
+
if isinstance(item,int):
|
300 |
+
return self.gene_id_name_dict.get(self.gene_token_id_dict.get(item, np.nan), np.nan)
|
301 |
+
if isinstance(item,tuple):
|
302 |
+
return tuple([self.gene_id_name_dict.get(self.gene_token_id_dict.get(i, np.nan), np.nan) for i in item])
|
geneformer/pretrainer.py
CHANGED
@@ -377,7 +377,7 @@ class GeneformerPreCollator(SpecialTokensMixin):
|
|
377 |
return_tensors = "tf" if return_tensors is None else return_tensors
|
378 |
elif is_torch_available() and _is_torch(first_element):
|
379 |
return_tensors = "pt" if return_tensors is None else return_tensors
|
380 |
-
|
381 |
return_tensors = "np" if return_tensors is None else return_tensors
|
382 |
else:
|
383 |
raise ValueError(
|
@@ -387,6 +387,7 @@ class GeneformerPreCollator(SpecialTokensMixin):
|
|
387 |
|
388 |
for key, value in encoded_inputs.items():
|
389 |
encoded_inputs[key] = to_py_obj(value)
|
|
|
390 |
|
391 |
# Convert padding_strategy in PaddingStrategy
|
392 |
padding_strategy, _, max_length, _ = self._get_padding_truncation_strategies(
|
|
|
377 |
return_tensors = "tf" if return_tensors is None else return_tensors
|
378 |
elif is_torch_available() and _is_torch(first_element):
|
379 |
return_tensors = "pt" if return_tensors is None else return_tensors
|
380 |
+
if isinstance(first_element, np.ndarray):
|
381 |
return_tensors = "np" if return_tensors is None else return_tensors
|
382 |
else:
|
383 |
raise ValueError(
|
|
|
387 |
|
388 |
for key, value in encoded_inputs.items():
|
389 |
encoded_inputs[key] = to_py_obj(value)
|
390 |
+
|
391 |
|
392 |
# Convert padding_strategy in PaddingStrategy
|
393 |
padding_strategy, _, max_length, _ = self._get_padding_truncation_strategies(
|