Oddp: Open Drug Design Pipeline

The Open Drug Design Pipeline project leverages advanced AI models to re-think an AI first approach to the drug discovery process aming for a design rather than a discovery pipeline, specifically targeting Antimicrobial Peptides (AMPs). This pipeline integrates state-of-the-art models across various stages, from initial sequence prediction to final protein design.

Note: This is a work in progress, some results in the are UI are mocked results from manual runs of the models.

Base Models

Model	Description
ProtTrans	A general protein language model (T5EncoderModel) for understanding protein sequences.
DiffDock	A diffusion generative model for molecular docking.
ProLLaMA	A large language model designed for multi-task protein language processing.

Pipeline Models

Model	Description
SignalP 6.0	Fine-tuned to predict all types of signal peptides from FASTA files, providing a classification probability for various pathways.
PeptideBERT	A language model for peptide property prediction, including hemolysis, solubility, and non-fouling properties.
DiffDock	Used for docking curated signal peptides with ligands of interest, generating complex 3D structures with confidence scores.
ProLLaMA	Generates novel proteins with specified functionalities based on user inputs.

Training Datasets

Dataset	Description
Unannotated protein sequences	Protein sequences across all domains of life.
DBAASPv3	Antimicrobial peptide activities.
PDBBind	Databases for docking purposes.
BindingMOAD	Databases for docking purposes.

Key Features

Feature	Description
Integration of multiple AI models	Streamlines the drug design process.
Utilization of large datasets	Ensures robust and accurate predictions.
Capability to design novel proteins	Tailored to specific functional requirements.
Web UI to run batch inferece	Web UI to manage inference over workloads

This AI-driven pipeline aims to significantly accelerate and enhance the precision of drug discovery, offering a powerful tool for developing new antimicrobial treatments.

Pipeline Models Fine-tuned for AMPs Discovery/Design

This pipeline incorporates foundational models such as ProtTrans, DiffDock, and ProLLaMA, each contributing unique capabilities to the drug design process. ProtTrans is a general protein language model designed for understanding and processing protein sequences. DiffDock is a diffusion generative model specialized in molecular docking, facilitating the prediction of protein-ligand interactions. ProLLaMA is a large language model tailored for multi-task protein language processing, capable of designing novel proteins with specific functionalities.

The AI First Drug Design Pipeline is structured into distinct steps, including signal peptide prediction, peptide property prediction, molecular docking, and de novo protein generation. Each step is fine-tuned to address specific aspects of AMP discovery and design, utilizing extensive datasets and advanced computational techniques to ensure high accuracy and efficiency

Pipeline Step 1: Predict SPs List

Field	Details
Model	SignalP 6.0
Description	Fine-tuned to predict all types of signal peptides from a FASTA file
Architecture	Fine-tuned for multi-classification
Base Model	T5EncoderModel
Torch Version	torch==1.13.1
CUDA Version	11.7
Training Dataset	Millions of unannotated protein sequences across all domains of life.
Paper	Nature Article
Input	Annotated FASTA file whole Genome e.g., Streptomyces JH010
Output	Classification probability: Sec/SPI, Sec/SPII, Tat/SPI, Tat/SPII, Sec/SPIII (general secretory or twin-arginine translocation (Tat) pathway). An annotated list in CSV with results.

Pipeline Step 2: Predict SPs Properties

Field	Details
Model	PeptideBERT
Description	Language Model for Peptide Property Prediction
Architecture	Fine-tuned for classification
Base Model	T5EncoderModel
Base Model	ProtTrans / ProtBERT
Torch Version	torch==2.0.1
Transformers	transformers==4.31.0
CUDA Version	11.7
Training Dataset	Database of Antimicrobial Activity and Structure of Peptides (DBAASPv3)
Hemolysis	9,316 sequences (19.6% positive (hemolytic), 80.4% negative (non-hemolytic))
Solubility	18,453 sequences (47.6% positive, 52.4% negative)
Nonfouling	3,600 sequences (20.9% positives, 79.1% negatives)
Paper	arXiv Article
Input	Amino Acid Sequence
Output	Classification probability: Hemolysis, Solubility, Non-fouling

Pipeline Step 3: Use the Curated SPs with Any Ligand of Interest

Field	Details
Model	DiffDock
Description	Diffusion generative model
Torch Version	torch==1.13.1
Python Version	python=3.9.18
CUDA Version	11.7
Training Dataset	PDBBind, BindingMOAD, DockGen, PoseBusters, van der Mers
Paper	arXiv Article
Input	Protein (.pdb), Ligand (SMILES string)
Output	Complex with 3D and confidence score

Pipeline Step 4: De Novo Protein Generation

Field	Details
Model	ProLLaMA
Description	Design novel proteins with desired functionalities
Architecture	LLM
Base Model	Meta LLama 2
Torch Version	torch==2.0.1
Transformers	transformers==4.35.0
CUDA Version	11.7
Paper	arXiv Article
Input	Amino Acid Sequence
Output	Generated proteins with specified functions based on the user's intent.

---

Setup

Requirements

Installation

Usage

Web UI

API

Training Pipeline

Models

Data