alpayariyak / bert-smiles-autocompletion-api

Fine-Tuning and Deploying BERT to predict the next element in a SMILES (Simplified Molecular Input Line Entry System) string.

BERT SMILES Autocompletion + API

BERT SMILES Autocompletion + API is a project fine-tuning and deploying a BERT model to predict the next element and character in a SMILES (Simplified Molecular Input Line Entry System) string. The API allows users to autocomplete SMILES strings with high accuracy, making it easier to access molecules without using drawing software.

Web App Demo of SMILES Autocompletion API

Table of Contents

• Fine-Tuned BERT Model for SMILES Autocompletion
  • Algorithm for SMILES Autocompletion with BERT
  • Advantages of using Model over Database Search
• BERT SMILES Autocompletion API
  • Installation
  • Usage
    • Endpoints
    • Query Parameters

Fine-Tuned BERT Model for SMILES Autocompletion

The BERT model was fine-tuned using a dataset of valid SMILES strings. Additional generated SMILES strings were also generated using the RDKit library. The dataset was preprocessed to create masked language model (MLM) training examples, where a portion of the SMILES strings was masked with a special [MASK] token. The objective of the MLM task is to predict the masked tokens based on the context provided by the surrounding unmasked tokens.

During the fine-tuning process, the model learned the syntactic and semantic patterns within the SMILES strings, enabling it to generate chemically valid suggestions for the masked positions.

Algorithm for SMILES Autocompletion with BERT

Algorithm for SMILES Autocompletion.

Advantages of using Model over Database Search

• Expanded Chemical Space: The model can generate exponentially more valid SMILES strings based on learned patterns, enabling exploration of novel and unexplored chemical structures.
• Robustness and Flexibility: The model adapts to different input SMILES strings and generates contextually appropriate suggestions, leading to more accurate and diverse results.
• Reduced Dependency on Database Size and Quality: By leveraging the model's learning capabilities, dependency on databases is minimized, making the autocompletion process more efficient and scalable.

BERT SMILES Autocompletion API

Installation

To set up and run the BERT SMILES Autocompletion API, follow these steps:

1. Clone the repository:

   $ git clone https://github.com/alpayariyak/BERT-SMILES-Autocompletion-API.git 
   $ cd BERT-SMILES-Autocompletion-API
   

2. Install the required packages:

   $ pip install -r requirements.txt
   

3. Run the Flask app:

   $ python autocompletionAPI.py
   

The API will be accessible at http://localhost:5000.

Usage

Endpoints

/autocomplete: autocompletes a given SMILES string using the fine-tuned BERT model, the database search, or both.

Query Parameters

smiles: The SMILES string to autocomplete. (required)

n_max_suggestions: The maximum number of suggestions to return (default: 5).

use_model: Set to true to use the BERT model for autocompletion (default: true).

use_database: Set to true to use the database search for autocompletion (default: true).

max_search_length: The maximum depth to search when using the BERT model (default: 10).