prot_bert_bfd-disoanno

This model is a fine-tuned version of Rostlab/prot_bert_bfd on the None dataset. It achieves the following results on the evaluation set:

• Loss: 0.4253
• Precision: 0.8161
• Recall: 0.8234
• F1: 0.8146

Model description

This is a token classification model designed to predict the intrinsically disordered regions of amino acid sequences.

Intended uses & limitations

This model works on amino acid sequences that are spaced between characters.

'0': No disorder

'1': Disordered

Example Inputs :

M S A I Q N L H S F D P F A D A S K G D D L L P A G T E D Y I H I R I Q Q R N G R K T L T T V Q G I A D D Y D K K K L V K A F K K K F A C N G T V I E H P E Y G E V I Q L Q G D Q R K N I C Q F L V E I G L A K D D Q L K V H G F

M A S R Q N N K Q E L D E R A R Q G E T V V P G G T G G K S L E A Q Q H L A E G R S K G G Q T R K E Q L G T E G Y Q E M G R K G G L S T V E K S G E E R A Q E E G I G I D E S K F R T G N N K N Q N Q N E D Q D K

M P P I A T R R G Q Y E P K V Q Q A K L S P D T I P L N P A D K T K D P L A R A D S L H H H V E S D S Q E D D K A A E E P P L S R K R W Q N R T F R R K G R R Q A P Y K H K

S G S D G G V C P K I L K K C R R D S D C P G A C I C R G N G Y C G

M G G K W S K S S V V G W P T V R E R M R R A E P A A D G V G A A S R D L E K H G A I T S S N T A A T N A A C A W L E A Q E E E E V G F P V T P Q V P L R P M T Y K A A V D L S H F L K E K G G L E G L I H S Q R R Q D I L D L W I Y H T Q G Y F P D W Q N Y T P G P G V R Y P L T F G W C Y K L V P V E P D K V E E A N K G E N T S L L H P V S L H G M D D P E R E V L E W R F D S R L A F H H V A R E L H P E Y F K N C

M R Y T D S R K L T P E T D A N H K T A S P Q P I R R I S S Q T L L G P D G K L I I D H D G Q E Y L L R K T Q A G K L L L T K

Training and evaluation data

Training and evaluation data were retrieved from https://www.csuligroup.com/DeepDISOBind/#Materials (Accessed March 2022).

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 2e-05
• train_batch_size: 8
• eval_batch_size: 8
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• num_epochs: 3

Training results

Training Loss	Epoch	Step	Validation Loss	Precision	Recall	F1
0.4895	1.0	61	0.4415	0.8257	0.8317	0.8262
0.5881	2.0	122	0.4242	0.8124	0.8201	0.8119
0.562	3.0	183	0.4253	0.8161	0.8234	0.8146

Framework versions

• Transformers 4.21.3
• Pytorch 1.12.1+cu113
• Datasets 2.4.0
• Tokenizers 0.12.1