damlab / HIV_BERT

HIV_BERT model

Table of Contents

• Summary
• Model Description
• Intended Uses & Limitations
• How to Use
• Training Data
• Training Procedure
  • Preprocessing
  • Training
• Evaluation Results
• BibTeX Entry and Citation Info

Summary

The HIV-BERT model was trained as a refinement of the ProtBert-BFD model for HIV centric tasks. It was refined with whole viral genomes from the Los Alamos HIV Sequence Database . This pretraining is important for HIV related tasks as the original BFD database contains few viral proteins making it sub-optimal when used as the basis for transfer learning tasks. This model and other related HIV prediction tasks have been published (link).

Model Description

Like the original ProtBert-BFD model , this model encodes each amino acid as an individual token. This model was trained using Masked Language Modeling: a process in which a random set of tokens are masked with the model trained on their prediction. This model was trained using the damlab/hiv-flt dataset with 256 amino acid chunks and a 15% mask rate.

Intended Uses & Limitations

As a masked language model this tool can be used to predict expected mutations using a masking approach. This could be used to identify highly mutated sequences, sequencing artifacts, or other contexts. As a BERT model, this tool can also be used as the base for transfer learning. This pretrained model could be used as the base when developing HIV-specific classification tasks.

How to use

As this is a BERT-style Masked Language learner, it can be used to determine the most likely amino acid at a masked position.

from transformers import pipeline

unmasker = pipeline("fill-mask", model="damlab/HIV_FLT")

unmasker(f"C T R P N [MASK] N T R K S I R I Q R G P G R A F V T I G K I G N M R Q A H C")

[
  {
    "score": 0.9581968188285828,
    "token": 17,
    "token_str": "N",
    "sequence": "C T R P N N N T R K S I R I Q R G P G R A F V T I G K I G N M R Q A H C"
  },
  {
    "score": 0.022986575961112976,
    "token": 12,
    "token_str": "K",
    "sequence": "C T R P N K N T R K S I R I Q R G P G R A F V T I G K I G N M R Q A H C"
  },
  {
    "score": 0.003997281193733215,
    "token": 14,
    "token_str": "D",
    "sequence": "C T R P N D N T R K S I R I Q R G P G R A F V T I G K I G N M R Q A H C"
  },
  {
    "score": 0.003636382520198822,
    "token": 15,
    "token_str": "T",
    "sequence": "C T R P N T N T R K S I R I Q R G P G R A F V T I G K I G N M R Q A H C"
  },
  {
    "score": 0.002701344434171915,
    "token": 10,
    "token_str": "S",
    "sequence": "C T R P N S N T R K S I R I Q R G P G R A F V T I G K I G N M R Q A H C"
  }
]

Training Data

The dataset damlab/HIV_FLT was used to refine the original rostlab/Prot-bert-bfd . This dataset contains 1790 full HIV genomes from across the globe. When translated, these genomes contain approximately 3.9 million amino-acid tokens.

Training Procedure

Preprocessing

As with the rostlab/Prot-bert-bfd model, the rare amino acids U, Z, O, and B were converted to X and spaces were added between each amino acid. All strings were concatenated and chunked into 256 token chunks for training. A random 20% of chunks were held for validation.

Training

Training was performed with the HuggingFace training module using the MaskedLM data loader with a 15% masking rate. The learning rate was set at E-5, 50K warm-up steps, and a cosine_with_restarts learning rate schedule and continued until 3 consecutive epochs did not improve the loss on the held-out dataset.

BibTeX Entry and Citation Info

[More Information Needed]