README.md

metadata

license: apache-2.0
base_model: nferruz/ProtGPT2
tags:
  - generated_from_trainer
metrics:
  - accuracy
model-index:
  - name: AmpGPT2
    results: []

AmpGPT2

AmpGPT2 is a language model capable of generating de novo antimicrobial peptides (AMPs). Over 95% of sequences generated by AmpGPT2 are predicted to have antimicrobial activities.

Model description

AmpGPT2 is a fine-tuned version of nferruz/ProtGPT2 based on the GPT2 Transformer architecture.

Model	sequences generated	AMP percentage (AMP%)	average length
AmpGPT2	1000	95.86	64.08
ProtGPT2	1000	51.85	222.59

The results demonstrate that AmpGPT2 outperformes ProtGPT2 in AMP%, suggesting the model learned from the AMP-specific data.
To validate the results the Antimicrobial Peptide Scanner vr.2 (https://www.dveltri.com/ascan/v2/ascan.html) was used, which is a deep learning tool specifically designed for AMP recognition.

Training and evaluation data

AmpGPT2 was trained using 32014 AMP sequences from the Compass (https://compass.mathematik.uni-marburg.de/) database.

How to use AmpGPT2

The example code below contains the ideal generation settings found while testing. The 'num_return_sequences' parameter specifies the amount of sequences generated. When generating more than 100 sequences at the same time, I recommend doing it in batches. The results can then be checked with the peptide scanner.

from transformers import pipeline
from transformers import GPT2LMHeadModel, GPT2Tokenizer

ampgpt2 = pipeline('text-generation', model="wabu/AmpGPT2")

model_amp = GPT2LMHeadModel.from_pretrained('wabu/AmpGPT2')
tokenizer_amp = GPT2Tokenizer.from_pretrained('wabu/AmpGPT2')

amp_sequences = ampgpt2( "", do_sample=True, repetition_penalty=1.2, num_return_sequences=10, eos_token_id=0 )

for i, seq in enumerate(amp_sequences):
    sequence_identifier = f"Sequence_{i + 1}"
    sequence = seq['generated_text'].replace('','').strip()

    print(f">{sequence_identifier}\n{sequence}")

Training hyperparameters and results

The following hyperparameters were used during training:

• learning_rate: 1e-05
• train_batch_size: 32
• eval_batch_size: 32
• seed: 42
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• num_epochs: 50.0

Training Loss	Epoch	Validation Loss	Accuracy
3.7948	50.0	3.9890	0.4213

Framework versions

• Transformers 4.38.0.dev0
• Pytorch 2.2.0+cu121
• Datasets 2.16.1
• Tokenizers 0.15.0

The model was trained on four NVIDIA A100 GPUs.