Commit
•
39ab62e
1
Parent(s):
4ce81ab
Fixed bugs related to overexpressing genes (#229)
Browse files- Fixed bugs related to overexpressing genes (72c65013e2fab0e3d97757f3a0af55b09a4864c7)
Co-authored-by: David Wen <[email protected]>
geneformer/in_silico_perturber.py
CHANGED
@@ -151,6 +151,7 @@ def overexpress_tokens(example):
|
|
151 |
if example["perturb_index"] != [-100]:
|
152 |
example = delete_indices(example)
|
153 |
[example["input_ids"].insert(0, token) for token in example["tokens_to_perturb"][::-1]]
|
|
|
154 |
return example
|
155 |
|
156 |
def remove_indices_from_emb(emb, indices_to_remove, gene_dim):
|
@@ -163,8 +164,8 @@ def remove_indices_from_emb(emb, indices_to_remove, gene_dim):
|
|
163 |
|
164 |
def remove_indices_from_emb_batch(emb_batch, list_of_indices_to_remove, gene_dim):
|
165 |
output_batch = torch.stack([
|
166 |
-
remove_indices_from_emb(emb_batch[i, :, :],
|
167 |
-
i,
|
168 |
])
|
169 |
return output_batch
|
170 |
|
@@ -179,7 +180,7 @@ def make_perturbation_batch(example_cell,
|
|
179 |
range_start = 1
|
180 |
elif perturb_type in ["delete","inhibit"]:
|
181 |
range_start = 0
|
182 |
-
indices_to_perturb = [[i] for i in range(range_start,example_cell["length"][0])]
|
183 |
elif combo_lvl>0 and (anchor_token is not None):
|
184 |
example_input_ids = example_cell["input_ids "][0]
|
185 |
anchor_index = example_input_ids.index(anchor_token[0])
|
@@ -323,47 +324,52 @@ def quant_cos_sims(model,
|
|
323 |
nproc):
|
324 |
cos = torch.nn.CosineSimilarity(dim=2)
|
325 |
total_batch_length = len(perturbation_batch)
|
|
|
326 |
if ((total_batch_length-1)/forward_batch_size).is_integer():
|
327 |
forward_batch_size = forward_batch_size-1
|
|
|
|
|
|
|
|
|
328 |
if cell_states_to_model is None:
|
329 |
-
if perturb_group == False: # (if perturb_group is True, original_emb is filtered_input_data)
|
330 |
-
comparison_batch = make_comparison_batch(original_emb, indices_to_perturb, perturb_group)
|
331 |
cos_sims = []
|
332 |
else:
|
333 |
possible_states = get_possible_states(cell_states_to_model)
|
334 |
-
cos_sims_vs_alt_dict = dict(zip(possible_states,[[] for
|
335 |
|
336 |
# measure length of each element in perturbation_batch
|
337 |
perturbation_batch = perturbation_batch.map(
|
338 |
measure_length, num_proc=nproc
|
339 |
)
|
340 |
-
|
341 |
-
|
342 |
-
|
343 |
-
|
344 |
-
|
345 |
-
|
346 |
-
minibatch_length_set
|
347 |
-
minibatch_lengths = perturbation_minibatch["length"]
|
348 |
-
if (len(minibatch_length_set) > 1) or (max(minibatch_length_set) > model_input_size):
|
349 |
needs_pad_or_trunc = True
|
350 |
else:
|
351 |
needs_pad_or_trunc = False
|
352 |
max_len = max(minibatch_length_set)
|
353 |
|
354 |
-
|
355 |
-
|
|
|
|
|
|
|
|
|
356 |
def pad_or_trunc_example(example):
|
357 |
example["input_ids"] = pad_or_truncate_encoding(example["input_ids"],
|
358 |
pad_token_id,
|
359 |
max_len)
|
360 |
return example
|
361 |
-
|
362 |
|
363 |
-
|
364 |
|
365 |
-
input_data_minibatch =
|
366 |
-
attention_mask = gen_attention_mask(
|
367 |
|
368 |
# extract embeddings for perturbation minibatch
|
369 |
with torch.no_grad():
|
@@ -371,9 +377,13 @@ def quant_cos_sims(model,
|
|
371 |
input_ids = input_data_minibatch.to("cuda"),
|
372 |
attention_mask = attention_mask
|
373 |
)
|
374 |
-
|
375 |
-
|
376 |
-
|
|
|
|
|
|
|
|
|
377 |
|
378 |
if len(indices_to_perturb)>1:
|
379 |
minibatch_emb = torch.squeeze(outputs.hidden_states[layer_to_quant])
|
@@ -386,7 +396,8 @@ def quant_cos_sims(model,
|
|
386 |
overexpressed_to_remove = 1
|
387 |
if perturb_group == True:
|
388 |
overexpressed_to_remove = len(tokens_to_perturb)
|
389 |
-
minibatch_emb = minibatch_emb[:,overexpressed_to_remove
|
|
|
390 |
|
391 |
# if quantifying single perturbation in multiple different cells, pad original batch and extract embs
|
392 |
if perturb_group == True:
|
@@ -394,56 +405,50 @@ def quant_cos_sims(model,
|
|
394 |
# truncate to the (model input size - # tokens to overexpress) to ensure comparability
|
395 |
# since max input size of perturb batch will be reduced by # tokens to overexpress
|
396 |
original_minibatch = original_emb.select([i for i in range(i, max_range)])
|
397 |
-
|
398 |
-
original_minibatch_length_set = set(original_minibatch["length"])
|
399 |
-
|
400 |
-
indices_to_perturb_minibatch = indices_to_perturb[i:i+forward_batch_size]
|
401 |
-
|
402 |
-
if perturb_type == "overexpress":
|
403 |
-
new_max_len = model_input_size - len(tokens_to_perturb)
|
404 |
-
else:
|
405 |
-
new_max_len = model_input_size
|
406 |
-
if (len(original_minibatch_length_set) > 1) or (max(original_minibatch_length_set) > new_max_len):
|
407 |
-
new_max_len = min(max(original_minibatch_length_set),new_max_len)
|
408 |
-
def pad_or_trunc_example(example):
|
409 |
-
example["input_ids"] = pad_or_truncate_encoding(example["input_ids"], pad_token_id, new_max_len)
|
410 |
-
return example
|
411 |
-
original_minibatch = original_minibatch.map(pad_or_trunc_example, num_proc=nproc)
|
412 |
-
original_minibatch.set_format(type="torch")
|
413 |
-
original_input_data_minibatch = original_minibatch["input_ids"]
|
414 |
-
attention_mask = gen_attention_mask(original_minibatch, new_max_len)
|
415 |
-
# extract embeddings for original minibatch
|
416 |
-
with torch.no_grad():
|
417 |
-
original_outputs = model(
|
418 |
-
input_ids = original_input_data_minibatch.to("cuda"),
|
419 |
-
attention_mask = attention_mask
|
420 |
-
)
|
421 |
-
del original_input_data_minibatch
|
422 |
-
del original_minibatch
|
423 |
-
del attention_mask
|
424 |
|
425 |
if len(indices_to_perturb)>1:
|
426 |
original_minibatch_emb = torch.squeeze(original_outputs.hidden_states[layer_to_quant])
|
427 |
else:
|
428 |
original_minibatch_emb = original_outputs.hidden_states[layer_to_quant]
|
429 |
|
430 |
-
#
|
431 |
-
|
432 |
-
#
|
433 |
-
if
|
434 |
-
|
435 |
-
|
436 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
437 |
|
|
|
|
|
|
|
|
|
438 |
# cosine similarity between original emb and batch items
|
439 |
if cell_states_to_model is None:
|
440 |
if perturb_group == False:
|
441 |
minibatch_comparison = comparison_batch[i:max_range]
|
442 |
elif perturb_group == True:
|
443 |
minibatch_comparison = original_minibatch_emb
|
444 |
-
|
445 |
cos_sims += [cos(minibatch_emb, minibatch_comparison).to("cpu")]
|
446 |
elif cell_states_to_model is not None:
|
|
|
|
|
|
|
|
|
|
|
447 |
for state in possible_states:
|
448 |
if perturb_group == False:
|
449 |
cos_sims_vs_alt_dict[state] += cos_sim_shift(original_emb,
|
@@ -455,12 +460,14 @@ def quant_cos_sims(model,
|
|
455 |
minibatch_emb,
|
456 |
state_embs_dict[state],
|
457 |
perturb_group,
|
458 |
-
|
459 |
-
|
460 |
del outputs
|
461 |
del minibatch_emb
|
462 |
if cell_states_to_model is None:
|
463 |
del minibatch_comparison
|
|
|
|
|
464 |
torch.cuda.empty_cache()
|
465 |
if cell_states_to_model is None:
|
466 |
cos_sims_stack = torch.cat(cos_sims)
|
@@ -470,6 +477,7 @@ def quant_cos_sims(model,
|
|
470 |
cos_sims_vs_alt_dict[state] = torch.cat(cos_sims_vs_alt_dict[state])
|
471 |
return cos_sims_vs_alt_dict
|
472 |
|
|
|
473 |
# calculate cos sim shift of perturbation with respect to origin and alternative cell
|
474 |
def cos_sim_shift(original_emb,
|
475 |
minibatch_emb,
|
@@ -478,34 +486,32 @@ def cos_sim_shift(original_emb,
|
|
478 |
original_minibatch_lengths = None,
|
479 |
minibatch_lengths = None):
|
480 |
cos = torch.nn.CosineSimilarity(dim=2)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
481 |
if not perturb_group:
|
482 |
-
original_emb = torch.mean(original_emb,dim=
|
483 |
-
|
484 |
-
origin_v_end = torch.squeeze(cos(original_emb, end_emb)) #test
|
485 |
else:
|
486 |
-
if original_emb.size() != minibatch_emb.size():
|
487 |
-
logger.error(
|
488 |
-
f"Embeddings are not the same dimensions. " \
|
489 |
-
f"original_emb is {original_emb.size()}. " \
|
490 |
-
f"minibatch_emb is {minibatch_emb.size()}. "
|
491 |
-
)
|
492 |
-
raise
|
493 |
-
|
494 |
if original_minibatch_lengths is not None:
|
495 |
original_emb = mean_nonpadding_embs(original_emb, original_minibatch_lengths)
|
496 |
# else:
|
497 |
# original_emb = torch.mean(original_emb,dim=1,keepdim=True)
|
498 |
|
499 |
end_emb = torch.unsqueeze(end_emb, 1)
|
500 |
-
origin_v_end = cos(original_emb, end_emb)
|
501 |
-
origin_v_end = torch.squeeze(origin_v_end)
|
502 |
if minibatch_lengths is not None:
|
503 |
perturb_emb = mean_nonpadding_embs(minibatch_emb, minibatch_lengths)
|
504 |
else:
|
505 |
perturb_emb = torch.mean(minibatch_emb,dim=1,keepdim=True)
|
506 |
-
|
507 |
perturb_v_end = cos(perturb_emb, end_emb)
|
508 |
perturb_v_end = torch.squeeze(perturb_v_end)
|
|
|
|
|
509 |
return [(perturb_v_end-origin_v_end).to("cpu")]
|
510 |
|
511 |
def pad_list(input_ids, pad_token_id, max_len):
|
@@ -1152,7 +1158,11 @@ class InSilicoPerturber:
|
|
1152 |
j_index = torch.squeeze(j_index)
|
1153 |
else:
|
1154 |
j_index = torch.tensor([j])
|
1155 |
-
|
|
|
|
|
|
|
|
|
1156 |
|
1157 |
if perturbed_gene.shape[0]==1:
|
1158 |
perturbed_gene = perturbed_gene.item()
|
@@ -1183,7 +1193,11 @@ class InSilicoPerturber:
|
|
1183 |
j_index = torch.squeeze(j_index)
|
1184 |
else:
|
1185 |
j_index = torch.tensor([j])
|
1186 |
-
|
|
|
|
|
|
|
|
|
1187 |
|
1188 |
if perturbed_gene.shape[0]==1:
|
1189 |
perturbed_gene = perturbed_gene.item()
|
|
|
151 |
if example["perturb_index"] != [-100]:
|
152 |
example = delete_indices(example)
|
153 |
[example["input_ids"].insert(0, token) for token in example["tokens_to_perturb"][::-1]]
|
154 |
+
|
155 |
return example
|
156 |
|
157 |
def remove_indices_from_emb(emb, indices_to_remove, gene_dim):
|
|
|
164 |
|
165 |
def remove_indices_from_emb_batch(emb_batch, list_of_indices_to_remove, gene_dim):
|
166 |
output_batch = torch.stack([
|
167 |
+
remove_indices_from_emb(emb_batch[i, :, :], idxs, gene_dim-1) for
|
168 |
+
i, idxs in enumerate(list_of_indices_to_remove)
|
169 |
])
|
170 |
return output_batch
|
171 |
|
|
|
180 |
range_start = 1
|
181 |
elif perturb_type in ["delete","inhibit"]:
|
182 |
range_start = 0
|
183 |
+
indices_to_perturb = [[i] for i in range(range_start, example_cell["length"][0])]
|
184 |
elif combo_lvl>0 and (anchor_token is not None):
|
185 |
example_input_ids = example_cell["input_ids "][0]
|
186 |
anchor_index = example_input_ids.index(anchor_token[0])
|
|
|
324 |
nproc):
|
325 |
cos = torch.nn.CosineSimilarity(dim=2)
|
326 |
total_batch_length = len(perturbation_batch)
|
327 |
+
|
328 |
if ((total_batch_length-1)/forward_batch_size).is_integer():
|
329 |
forward_batch_size = forward_batch_size-1
|
330 |
+
|
331 |
+
if perturb_group == False:
|
332 |
+
comparison_batch = make_comparison_batch(original_emb, indices_to_perturb, perturb_group)
|
333 |
+
|
334 |
if cell_states_to_model is None:
|
|
|
|
|
335 |
cos_sims = []
|
336 |
else:
|
337 |
possible_states = get_possible_states(cell_states_to_model)
|
338 |
+
cos_sims_vs_alt_dict = dict(zip(possible_states,[[] for _ in range(len(possible_states))]))
|
339 |
|
340 |
# measure length of each element in perturbation_batch
|
341 |
perturbation_batch = perturbation_batch.map(
|
342 |
measure_length, num_proc=nproc
|
343 |
)
|
344 |
+
|
345 |
+
def compute_batch_embeddings(minibatch, _max_len = None):
|
346 |
+
minibatch_lengths = minibatch["length"]
|
347 |
+
minibatch_length_set = set(minibatch_lengths)
|
348 |
+
max_len = model_input_size
|
349 |
+
|
350 |
+
if (len(minibatch_length_set) > 1) or (max(minibatch_length_set) > max_len):
|
|
|
|
|
351 |
needs_pad_or_trunc = True
|
352 |
else:
|
353 |
needs_pad_or_trunc = False
|
354 |
max_len = max(minibatch_length_set)
|
355 |
|
356 |
+
|
357 |
+
if needs_pad_or_trunc == True:
|
358 |
+
if _max_len is None:
|
359 |
+
max_len = min(max(minibatch_length_set), max_len)
|
360 |
+
else:
|
361 |
+
max_len = _max_len
|
362 |
def pad_or_trunc_example(example):
|
363 |
example["input_ids"] = pad_or_truncate_encoding(example["input_ids"],
|
364 |
pad_token_id,
|
365 |
max_len)
|
366 |
return example
|
367 |
+
minibatch = minibatch.map(pad_or_trunc_example, num_proc=nproc)
|
368 |
|
369 |
+
minibatch.set_format(type="torch")
|
370 |
|
371 |
+
input_data_minibatch = minibatch["input_ids"]
|
372 |
+
attention_mask = gen_attention_mask(minibatch, max_len)
|
373 |
|
374 |
# extract embeddings for perturbation minibatch
|
375 |
with torch.no_grad():
|
|
|
377 |
input_ids = input_data_minibatch.to("cuda"),
|
378 |
attention_mask = attention_mask
|
379 |
)
|
380 |
+
|
381 |
+
return outputs, max_len
|
382 |
+
|
383 |
+
for i in range(0, total_batch_length, forward_batch_size):
|
384 |
+
max_range = min(i+forward_batch_size, total_batch_length)
|
385 |
+
perturbation_minibatch = perturbation_batch.select([i for i in range(i, max_range)])
|
386 |
+
outputs, mini_max_len = compute_batch_embeddings(perturbation_minibatch)
|
387 |
|
388 |
if len(indices_to_perturb)>1:
|
389 |
minibatch_emb = torch.squeeze(outputs.hidden_states[layer_to_quant])
|
|
|
396 |
overexpressed_to_remove = 1
|
397 |
if perturb_group == True:
|
398 |
overexpressed_to_remove = len(tokens_to_perturb)
|
399 |
+
minibatch_emb = minibatch_emb[:, overexpressed_to_remove: ,:]
|
400 |
+
|
401 |
|
402 |
# if quantifying single perturbation in multiple different cells, pad original batch and extract embs
|
403 |
if perturb_group == True:
|
|
|
405 |
# truncate to the (model input size - # tokens to overexpress) to ensure comparability
|
406 |
# since max input size of perturb batch will be reduced by # tokens to overexpress
|
407 |
original_minibatch = original_emb.select([i for i in range(i, max_range)])
|
408 |
+
original_outputs, orig_max_len = compute_batch_embeddings(original_minibatch, mini_max_len)
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
409 |
|
410 |
if len(indices_to_perturb)>1:
|
411 |
original_minibatch_emb = torch.squeeze(original_outputs.hidden_states[layer_to_quant])
|
412 |
else:
|
413 |
original_minibatch_emb = original_outputs.hidden_states[layer_to_quant]
|
414 |
|
415 |
+
# if we overexpress genes that aren't already expressed,
|
416 |
+
# we need to remove genes to make sure the embeddings are of a consistent size
|
417 |
+
# get rid of the bottom n genes/padding since those will get truncated anyways
|
418 |
+
# multiple perturbations is more complicated because if 1 out of n perturbed genes is expressed
|
419 |
+
# the idxs will still not be [-100]
|
420 |
+
if len(tokens_to_perturb) == 1:
|
421 |
+
indices_to_perturb_minibatch = [idx if idx != [-100] else [orig_max_len - 1]
|
422 |
+
for idx in indices_to_perturb[i:max_range]]
|
423 |
+
else:
|
424 |
+
num_perturbed = len(tokens_to_perturb)
|
425 |
+
indices_to_perturb_minibatch = []
|
426 |
+
end_range = [i for i in range(orig_max_len - tokens_to_perturb, orig_max_len)]
|
427 |
+
for idx in indices_to_perturb[i:i+max_range]:
|
428 |
+
if idx == [-100]:
|
429 |
+
indices_to_perturb_minibatch.append(end_range)
|
430 |
+
elif len(idx) < len(tokens_to_perturb):
|
431 |
+
indices_to_perturb_minibatch.append(idx + end_range[-num_perturbed:])
|
432 |
+
else:
|
433 |
+
indices_to_perturb_minibatch.append(idx)
|
434 |
|
435 |
+
original_minibatch_emb = remove_indices_from_emb_batch(original_minibatch_emb,
|
436 |
+
indices_to_perturb_minibatch,
|
437 |
+
gene_dim=1)
|
438 |
+
|
439 |
# cosine similarity between original emb and batch items
|
440 |
if cell_states_to_model is None:
|
441 |
if perturb_group == False:
|
442 |
minibatch_comparison = comparison_batch[i:max_range]
|
443 |
elif perturb_group == True:
|
444 |
minibatch_comparison = original_minibatch_emb
|
|
|
445 |
cos_sims += [cos(minibatch_emb, minibatch_comparison).to("cpu")]
|
446 |
elif cell_states_to_model is not None:
|
447 |
+
if perturb_group == False:
|
448 |
+
original_emb = comparison_batch[i:max_range]
|
449 |
+
else:
|
450 |
+
original_minibatch_lengths = torch.tensor(original_minibatch["length"], device="cuda")
|
451 |
+
minibatch_lengths = torch.tensor(perturbation_minibatch["length"], device="cuda")
|
452 |
for state in possible_states:
|
453 |
if perturb_group == False:
|
454 |
cos_sims_vs_alt_dict[state] += cos_sim_shift(original_emb,
|
|
|
460 |
minibatch_emb,
|
461 |
state_embs_dict[state],
|
462 |
perturb_group,
|
463 |
+
original_minibatch_lengths,
|
464 |
+
minibatch_lengths)
|
465 |
del outputs
|
466 |
del minibatch_emb
|
467 |
if cell_states_to_model is None:
|
468 |
del minibatch_comparison
|
469 |
+
if perturb_group == True:
|
470 |
+
del original_minibatch_emb
|
471 |
torch.cuda.empty_cache()
|
472 |
if cell_states_to_model is None:
|
473 |
cos_sims_stack = torch.cat(cos_sims)
|
|
|
477 |
cos_sims_vs_alt_dict[state] = torch.cat(cos_sims_vs_alt_dict[state])
|
478 |
return cos_sims_vs_alt_dict
|
479 |
|
480 |
+
|
481 |
# calculate cos sim shift of perturbation with respect to origin and alternative cell
|
482 |
def cos_sim_shift(original_emb,
|
483 |
minibatch_emb,
|
|
|
486 |
original_minibatch_lengths = None,
|
487 |
minibatch_lengths = None):
|
488 |
cos = torch.nn.CosineSimilarity(dim=2)
|
489 |
+
if original_emb.size() != minibatch_emb.size():
|
490 |
+
logger.error(
|
491 |
+
f"Embeddings are not the same dimensions. " \
|
492 |
+
f"original_emb is {original_emb.size()}. " \
|
493 |
+
f"minibatch_emb is {minibatch_emb.size()}. "
|
494 |
+
)
|
495 |
+
raise
|
496 |
if not perturb_group:
|
497 |
+
original_emb = torch.mean(original_emb,dim=1,keepdim=True)
|
498 |
+
origin_v_end = torch.squeeze(cos(original_emb, end_emb))
|
|
|
499 |
else:
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
500 |
if original_minibatch_lengths is not None:
|
501 |
original_emb = mean_nonpadding_embs(original_emb, original_minibatch_lengths)
|
502 |
# else:
|
503 |
# original_emb = torch.mean(original_emb,dim=1,keepdim=True)
|
504 |
|
505 |
end_emb = torch.unsqueeze(end_emb, 1)
|
506 |
+
origin_v_end = torch.squeeze(cos(original_emb, end_emb))
|
|
|
507 |
if minibatch_lengths is not None:
|
508 |
perturb_emb = mean_nonpadding_embs(minibatch_emb, minibatch_lengths)
|
509 |
else:
|
510 |
perturb_emb = torch.mean(minibatch_emb,dim=1,keepdim=True)
|
|
|
511 |
perturb_v_end = cos(perturb_emb, end_emb)
|
512 |
perturb_v_end = torch.squeeze(perturb_v_end)
|
513 |
+
if (perturb_v_end-origin_v_end).numel() == 1:
|
514 |
+
return [([perturb_v_end-origin_v_end]).to("cpu")]
|
515 |
return [(perturb_v_end-origin_v_end).to("cpu")]
|
516 |
|
517 |
def pad_list(input_ids, pad_token_id, max_len):
|
|
|
1158 |
j_index = torch.squeeze(j_index)
|
1159 |
else:
|
1160 |
j_index = torch.tensor([j])
|
1161 |
+
|
1162 |
+
if self.perturb_type in ("overexpress", "activate"):
|
1163 |
+
perturbed_gene = torch.index_select(gene_list, 0, j_index + 1)
|
1164 |
+
else:
|
1165 |
+
perturbed_gene = torch.index_select(gene_list, 0, j_index)
|
1166 |
|
1167 |
if perturbed_gene.shape[0]==1:
|
1168 |
perturbed_gene = perturbed_gene.item()
|
|
|
1193 |
j_index = torch.squeeze(j_index)
|
1194 |
else:
|
1195 |
j_index = torch.tensor([j])
|
1196 |
+
|
1197 |
+
if self.perturb_type in ("overexpress", "activate"):
|
1198 |
+
perturbed_gene = torch.index_select(gene_list, 0, j_index + 1)
|
1199 |
+
else:
|
1200 |
+
perturbed_gene = torch.index_select(gene_list, 0, j_index)
|
1201 |
|
1202 |
if perturbed_gene.shape[0]==1:
|
1203 |
perturbed_gene = perturbed_gene.item()
|