1

Antigen-aware antibody design with encoder-decoder architectures

Author

Georgii Sarapulov

​

Supervisor

Natalia Zenkova

BIOCAD

​

2

Introduction

De novo design of antibodies binding to specific antigens has an enormous significance for developing therapeutic treatment for various diseases.

​

Challenges:

• complex interface geometry
• flexible conformations of complementarity-determining regions (CDRs)
• existing structure prediction methods lack the sufficient accuracy for CDR loops

Source: https://opig.stats.ox.ac.uk/webapps/newsabdab/sabdab/about/

3

Goal:

Build machine learning model for Fv-fragment (VH+VL) amino acid sequence generation conditioned on the antigen with a known epitope 3D structure

​

Objectives:

1. Build a data processing and training pipeline
2. Develop a set of metrics measuring the performance of generative models
3. Perform experiments with various antigen representations and encoder-decoder architectures

Project goal and objectives

4

Epitope representation: amino acid sequence of linear epitope segment (includes most of the contact residues)

​

​

​

Model architecture: vanilla transformer encoder-decoder mapping linear epitope sequence to VH+VL sequence

​

Methods: sequence-to-sequence

Transformer architecture�(from Vaswani et al. - Attention Is All You Need)

<AG>GARGLTGG<EOS>

<VH>VKLV…TVSA

<VL>DIVM…TKLEIK

5

Epitope representation: graph with atom elements as node features and pairwise euclidean distances as edge features; edge exists only if the distance between atoms is less than 5Å

​

​

Model architecture: equivariant graph encoder (EGNN) + transformer decoder mapping epitope structure to VH+VL sequence

​

Methods: structure-to-sequence

<VH>VKLV…TVSA

<VL>DIVM…TKLEIK

EGNN encoder + Transformer decoder�(adapted from Vaswani et al. - Attention Is All You Need)

6

Methods: estimating the performance

Idea: align generated sequence to the native antibody from antibody-antigen complex, calculate the percentage of matched amino acids in different regions: FRs, CDRs, paratope

​

Example:

..X.XX.* * * ...... *

Reference: -VQLQQSGAELVK-PGASVKLSCTASGFNIKDTYMYWVKQRPEQGLEWIGRIDPANGDTKYDP...

|||.||||| || ||||||.||.|||......|||||.|.|.|||||.|||.||||||||..

Generated: QVQLVQSGAE-VKKPGASVKVSCKASGYTFTSYYMYWVRQAPGQGLEWMGRIGPANGDTKYAQ...

FR Score: 0.7755, CDR Score: 0.6818, Paratope Score: 0.7273

​

Notation:

. - CDR

X - Contact in CDR

* - Contact outside CDR

7

Results: human-like antibodies

> VH

QVQLVQGAEVKKPGASVKVSCKASGYTFTSYIIHWVRQAPGQGLEWMGWINPNSGGTSYAQKFQGRVTMTRDTSTSTAYMELSSLRSDDTAVYYCAREGSPFYFDYWGQGTLVTVSS

Close reference: IGHV1-2*02_IGHD1-1*01_IGHJ4*01_Human

Similarity: 0.898

​

> VL

DIQMTQSPSSLSASVGDRVTITCRASQSISSYLNWYQQKPGKAPKLLIYYASNLQSGVPSRFSGSGSGTDFTLTISSLQPEDFATYYCQQSYSTPPTFGQGTKVEIK

Close reference: IGKV1-39*01_IGKJ1*01_Human

Similarity: 0.972

Models generate antibody sequences very similar to human antibodies

Generated sample for epitope LDSFKEELDKYFKNH (from SARS-CoV-2 spike stem helix peptide) with CDRs highlighted:

​

​

8

Results: cross-attention patterns

Sequence-to-sequence model attends to different epitope amino acids when generating CDRs and almost ignores the antigen while generating FRs

​

Left: cross-attention weights between heavy chain CDR3 and epitope amino acids

Right: cross-attention weights between heavy chain FR4 and epitope amino acids for the same attention head, attention collapses to end-of-sequence token

​

​

9

Results: overall scores

Models show the ability to reconstruct amino acids of a native antibody both in CDRs and in positions involved in a contact with the epitope.

​

Table shows the % of matched amino acids between native and generated antibodies in CDRs and paratope averaged across the test set. Sequence-to-sequence model performs better so far

10

Conclusion and research plans

A more information-rich representation of epitope structure (beyond atom elements and pairwise distances) will most likely lead to an increased performance

A multimodal learning approach:

1. pretrain graph encoder on bigger subset of protein structures from PDB
2. pretrain sequence decoder on antibody sequences from OAS database
3. fine-tune the combined model on a downstream task with limited dataset

​

​

GitHub: https://github.com/norsage/a3d