evankomp/learn2therm

Purpose: classifies protein sequence into Thermophilic (>= 60C) or Mesophilic (<30C) by host organism growth temperature.

Usage: Prepare sequences identically to using the original pretrained model:

from transformers import BertModelForSequenceClassification, BertTokenizer
import torch
import re
tokenizer = BertTokenizer.from_pretrained("evankomp/learn2therm", do_lower_case=False )
model = BertModelForSequenceClassification.from_pretrained("evankomp/learn2therm")
sequence_Example = "A E T C Z A O"
sequence_Example = re.sub(r"[UZOB]", "X", sequence_Example)
encoded_input = tokenizer(sequence_Example, return_tensors='pt')
output = torch.argmax(model(**encoded_input), dim=1)

1 indicates thermophilic, 0 mesophilic.

Training: ProteinBERT (Rostlab/prot_bert) was fine tuned on a class balanced version of learn2therm (see here), about 250k protein amino acid sequences.

Training parameters below:

TrainingArguments(
_n_gpu=1,
adafactor=False,
adam_beta1=0.9,
adam_beta2=0.999,
adam_epsilon=1e-08,
auto_find_batch_size=False,
bf16=False,
bf16_full_eval=False,
data_seed=None,
dataloader_drop_last=False,
dataloader_num_workers=0,
dataloader_pin_memory=True,
ddp_bucket_cap_mb=None,
ddp_find_unused_parameters=None,
ddp_timeout=1800,
debug=[],
deepspeed=None,
disable_tqdm=False,
do_eval=True,
do_predict=False,
do_train=True,
eval_accumulation_steps=25,
eval_delay=0,
eval_steps=6,
evaluation_strategy=steps,
fp16=True,
fp16_backend=auto,
fp16_full_eval=False,
fp16_opt_level=O1,
fsdp=[],
fsdp_min_num_params=0,
fsdp_transformer_layer_cls_to_wrap=None,
full_determinism=False,
gradient_accumulation_steps=25,
gradient_checkpointing=True,
greater_is_better=False,
group_by_length=False,
half_precision_backend=cuda_amp,
hub_model_id=None,
hub_private_repo=False,
hub_strategy=every_save,
hub_token=<HUB_TOKEN>,
ignore_data_skip=False,
include_inputs_for_metrics=False,
jit_mode_eval=False,
label_names=None,
label_smoothing_factor=0.0,
learning_rate=5e-05,
length_column_name=length,
load_best_model_at_end=True,
local_rank=0,
log_level=info,
log_level_replica=passive,
log_on_each_node=True,
logging_dir=./data/ogt_protein_classifier/model/runs/Jun19_12-16-35_g3070,
logging_first_step=False,
logging_nan_inf_filter=True,
logging_steps=1,
logging_strategy=steps,
lr_scheduler_type=linear,
max_grad_norm=1.0,
max_steps=-1,
metric_for_best_model=loss,
mp_parameters=,
no_cuda=False,
num_train_epochs=2,
optim=adamw_hf,
optim_args=None,
output_dir=./data/ogt_protein_classifier/model,
overwrite_output_dir=False,
past_index=-1,
per_device_eval_batch_size=32,
per_device_train_batch_size=32,
prediction_loss_only=False,
push_to_hub=False,
push_to_hub_model_id=None,
push_to_hub_organization=None,
push_to_hub_token=<PUSH_TO_HUB_TOKEN>,
ray_scope=last,
remove_unused_columns=True,
report_to=['tensorboard', 'codecarbon'],
resume_from_checkpoint=None,
run_name=./data/ogt_protein_classifier/model,
save_on_each_node=False,
save_steps=6,
save_strategy=steps,
save_total_limit=None,
seed=42,
sharded_ddp=[],
skip_memory_metrics=True,
tf32=None,
torch_compile=False,
torch_compile_backend=None,
torch_compile_mode=None,
torchdynamo=None,
tpu_metrics_debug=False,
tpu_num_cores=None,
use_ipex=False,
use_legacy_prediction_loop=False,
use_mps_device=False,
warmup_ratio=0.0,
warmup_steps=0,
weight_decay=0.0,
xpu_backend=None,
)

See the training repository for code.