Use P2PXML Online
Commercially Available P2PXML No-Code Web Server
P2PXML: Deep Geometric Framework to Predict Antibody-Antigen Binding Affinity
P2PXML is a deep learning framework and pair of curated datasets for predicting antibody-antigen binding affinity from sequences, 3D structures or both. It was developed by Nuwan Bandara, Dasun Premathilaka, Sachini Chandanayake, Sahan Hettiarachchi, Vithurshan Varenthirarajah, Aravinda Munasinghe, Kaushalya Madhawa and Subodha Charles, with affiliations at the University of Moratuwa, Singapore Management University, the University of Colombo, Pfizer and the University of Tokyo. The preprint is licensed CC BY 4.0, and no software license is stated.
Binding affinity measures how strongly an antibody binds its antigen, which matters for antibody engineering and drug development. Physics-based methods such as docking and molecular dynamics are computationally expensive for large antibodies and antigens.
Existing deep learning predictors depend heavily on structure quality, often ignore evolutionary information, and are trained on datasets covering only one antigen variant. P2PXML tackles this with larger, more generalized datasets and a model that combines a structure branch with a sequence branch.
Item | Detail from the paper |
|---|---|
Datasets | P2PXML-Seq (FASTA sequences) and P2PXML-PDB (3D structures) |
Antigens covered | Including SARS-CoV-2, HIV, MERS and influenza |
Inputs | Antibody and antigen sequences and/or PDB structures (proteins only) |
Target value | Binding affinity based on IC50 (log10 IC50 in the main text) |
Best model | Combined-V2 with pre-trained sequence weights |
Speed | About 1 minute for 125 pairs on a GPU, about 3 minutes on a CPU |
How P2PXML Combines Sequence and Structure for Affinity Prediction
Data curation: public datasets were standardized, and repeated pairs were resolved using the median IC50.
Missing structures: structures were generated by homology modeling with MODELLER, or by AlphaFold-V2 multimer through ColabFold.
Structure model: antibody and antigen are atom graphs with edges between atoms within 5 A, processed by graph convolution and graph attention layers.
Sequence model: sequences are encoded with VHSE8, an 8-dimensional physicochemical encoding, and processed by a transformer with self-attention and cross-attention.
Combined model: cross-attention shares information between pathways, and the two predictions are averaged with weights 0.45 and 0.55.
P2PXML Results Versus State-of-the-Art Affinity Predictors
Metric | Reported result |
|---|---|
MAE on P2PXML-PDB test set | 1.0005, at least 10.6% better than compared state-of-the-art models |
Pearson r | 0.8703 |
Spearman rho | 0.9450 |
Sequence model versus VirusNet | About 4% MAE improvement |
Structure model versus state of the art | Marginal gain, about 0.08% in MAE |
Ablation Findings
Sharing information between the pathways improved MAE by about 5% (Combined-V1 versus the base model), and cross-attention improved results further. VHSE8 encoding beat one-hot and ProtT5 embeddings in the authors' tests.
What is Tamarind Bio?
Tamarind Bio is a no-code bioinformatics platform built to give life scientists and researchers access to powerful computational tools. Many cutting-edge machine learning models are hard to deploy and use. Tamarind provides an intuitive, web-based environment that removes the complexity of high-performance computing, software dependencies and command-line interfaces.
The platform is designed for biologists, chemists and other researchers who may not have a background in programming or cloud infrastructure but want to run models on their own data. Key features include:
A user-friendly graphical interface for setting up and launching experiments
A robust API for integration into existing research pipelines
An automated system for managing and scaling computational resources
Tamarind treats information and data security as a top priority, as detailed in its Trust Center and Terms of Service.
Accelerating Discovery with P2PXML on Tamarind Bio
Antibody engineering: estimate how strongly candidate antibodies bind a target antigen.
Drug development screening: prioritize antibody candidates faster than physics-based docking or molecular dynamics.
Viral antigen studies: predict affinity across antigens such as SARS-CoV-2, HIV, MERS and influenza.
Large-scale triage: score many antibody-antigen pairs in minutes.
How to Use P2PXML on Tamarind Bio
Open the tool: log in to tamarind.bio and open P2PXML.
Provide the antibody: supply its sequence (FASTA) or PDB structure.
Provide the antigen: supply its sequence or PDB structure, protein antigens only.
Select the model: choose combined, sequence-only or structure-only prediction, if exposed.
Run the job: submit and wait for processing.
Download outputs: retrieve the predicted binding affinity values.
Interpret and rank: compare predictions across variants, noting that unit conventions should be checked against the released materials.
Things to Keep in Mind
Many data points lack binding-site information, so antibody and antigen structures are treated separately rather than as a complex.
Structures are often computationally generated, so prediction quality depends on them.
Some IC50 values were approximated using assumptions such as non-competitive inhibition.
The structure model's gain over prior work is marginal.
Only protein antibodies and antigens are covered.
Source: Bandara N, Premathilaka D, Chandanayake S, Hettiarachchi S, Varenthirarajah V, Munasinghe A, Madhawa K, Charles S. Deep Geometric Framework to Predict Antibody-Antigen Binding Affinity. bioRxiv 2024. DOI: 10.1101/2024.06.09.598103. Project page: https://drug-discovery-entc.github.io/p2pxml/