Use P2PXML Online

P2PXML: Predict Antibody-Antigen Binding Affinity

Predict antibody-antigen binding affinity with state-of-the-art accuracy using a deep geometric framework that combines atomistic structures and evolutionary protein sequences.

• Estimated Run Time: Under 1 minute on GPU, ~3 minutes on CPU.

• Input Types: PDB Files (Structures) or Text Sequences (FASTA format).

• Output: Quantitative binding affinity (IC50 in µg/ml).

How P2PXML Works

P2PXML (Predicting Protein Interactions via XML Framework) is an end-to-end deep learning framework designed to circumvent the heavy computational complexity of traditional wet-lab assays, molecular docking, and molecular dynamics (MD) simulations. It models antibody-antigen binding kinetics with a unique dual-pipeline architecture:

• The Sequence-Based Pipeline: Processes the entire amino acid sequence using a Parallel Transformer architecture with cross-attention. Instead of relying on general deep learning-based embeddings, it leverages VHSE8 encoding (an 8D vector tracking Hydrophobic, Steric, and Electronic properties of amino acids), capturing fine-grained evolutionary protein variations upon mutation.

• The Structure-Based Pipeline: Translates 3D structural data into graphs where atoms serve as nodes and distances act as edges. It routes data through parallel Graph Convolutional Networks (GCN) and Graph Attention Networks (GATConv) to evaluate neighborhood feature aggregations and atomistic-level structural configurations.

• The Combined Information Pipeline: P2PXML leverages a hierarchal cross-attention information-sharing mechanism to capture cross-domain features from both sequences and structures. It produces an ensemble prediction optimized via a combined loss function, mirroring natural chemical binding potentials.

State-of-the-Art Performance Benchmarks

P2PXML delivers a 10.6% improvement in Mean Absolute Error (MAE) compared to existing state-of-the-art models. It is backed by a remarkably strong linear and rank correlation, sporting a Pearson correlation coefficient of 0.8703 and a Spearman's rank correlation coefficient of 0.9450 between true and predicted targets.

Scientific Breakthroughs in Antibody-Antigen Modeling

Broad Generalizability Across Pathogens

Most existing deep learning tools are over-fit to a single antigen variant or specific viral strain. P2PXML is built on the most extensive and generalized antibody-antigen datasets compiled in literature (P2PXML-Seq with 111,845 datapoints and P2PXML-PDB with 8,475 processed structural points). This ensures highly generalized performance across multiple highly mutable pathogenic targets, including:

• SARS-CoV-2

• Human Immunodeficiency Virus (HIV)

• MERS-CoV

• Influenza (Flu) Viruses

AI-Assisted Structure Resolution

Traditional binding analysis strictly depends on ultra-high-resolution, experimentally validated 3D crystalline structures. P2PXML addresses structural gaps through an integrated automated curation pipeline:

• Homology Modeling: Uses 'MODELLER' and MDAnalysis packages to construct atom-resolution structural variants when a template 3D coordinate and mutation profile are known.

• AlphaFold-V2 Multimer: Deploys a ColabFold and MMseqs2-driven homology search pipeline to predict atomistic-level structural coordinates when template structures are entirely absent.

What is Tamarind Bio?

Tamarind Bio is a professional, no-code computational platform built to democratize access to cutting-edge bioinformatics and drug discovery algorithms. Realizing that many leading machine learning tools are locked behind complex command-line interfaces, rigid software dependencies, and highly demanding cloud infrastructures, Tamarind Bio delivers a simple, web-based graphical interface.

By fully abstracting high-performance GPU orchestration and cloud parallelization, Tamarind Bio allows biologists, immunologists, and pharmaceutical chemists to effortlessly deploy high-throughput models on thousands of sequences or structures. Built with enterprise-grade, SOC 2 compliant data privacy and security measures, the platform guarantees total isolation for sensitive data—ensuring that you retain full ownership of your data, structural inputs, and molecular IP.

How to Use P2PXML on Tamarind Bio

Tamarind Bio abstracts away the underlying neural architecture, allowing you to quickly compute binding affinities via an intuitive web dashboard:

1. Access the Tool: Log in to your secure dashboard on the Tamarind Bio website and select P2PXML from the structural developability and binder analysis module.

2. Choose Your Input Format: Select your target entry parameters based on your available data:

   • PDB Structure Upload: Upload the .pdb files for individual antibody and antigen chains to trigger the dual sequence-structure model architecture.

   • Text Sequences (FASTA): Input the letter codes for the antibody and antigen protein sequences to utilize the standalone sequence-based model.

3. Configure Settings (Optional): If you upload raw sequences lacking crystal structures, choose whether to resolve structural gaps using the integrated Homology Modeling pipeline (requires a known template structure and mutation profile) or the AlphaFold-V2 Multimer engine.

4. Launch Inference: Click Submit. Tamarind Bio schedules and orchestrates the backend GPU pipeline. Results take less than 60 seconds on standard accelerated instances.

5. Analyze Outputs: Review your interactive reports directly on screen. P2PXML outputs a precise, quantitative measurement of binding affinity (IC50 in µg/ml) along with intermediate structural and evolutionary pathway embeddings. Download the clean data files for downstream therapeutic testing or affinity maturation steps.

Source