Protein inpainting to design better binding interfaces with ProtFill

What do we want to say:

• why why why

• problem: de novo design models are built mainly for antibodies

• problem: given a sequence / structure of a binder and a target, can we optimize the designs by redesigning both seq & struc simultaneously

• protfill is our model for protein inpainting

• protfill is trained on antibody antigen datasets and can recover antibody binding loops?

• this problem can be nicely represented by diffusion, first used in image gen, now applied to protein design:

to . Today we introduce ProtFill, a novel method for protein design. While existing alternatives exclusively focus on designing antibodies and nanobodies, ProtFill outperforms them in that task and extends to optimising binding in any other protein.

Architecture

The model consists of one shared encoder and two decoders (one for sequence and one for structure outputs). Both the encoder and the decoders consist of three stacked message passing layers. The encoder generates a structure embedding which is concatenated with a simple sequence embedding (the output of a single embedding layer) and passed to the decoders

ProtFill builds upon the foundation laid by previous models but takes it a step further. It is based on a novel neural network architecture called GVPe that allows passing a richer representation of the protein data along the model. ProtFill also boosts its performance by utilising the idea of recycling, or gradually refining the model predictions.

Diffusion has been shown to generate groundbreaking results for tasks like image generation, and in protein design it opens up the potential for generating molecules with fine-grained properties.

We have profill: oneshot model

We have protfilldiff: diffusion model there you can condition more finetune the generation

In training ProtFill, we've adopted a method known as diffusion, which involves carefully adding and removing noise from the data to teach the model about different protein structures and sequences.

What can you do with ProtFill

We designed protfil to redesign antibody antigen interfaces, model is trained to generate new CDRs on demand.

We show good sequence prediction (amino acid recovery)

We show good structure prediction (rmsd)

MEAN: codesign model, one shot

diffAb: codesign models, diffusion

retrained on the same data set

• we see quite improved sequence recovery while showing comparable structure prediction that are close to the ground

suited for antibody optimization

you go and you take a parental antibody e.g.

• from pdb

• from immunization, phage display etc and generate structure using antibody structure model

if you provide the model with a weakly binding antibody and an antigen epitope

you generate new CDRs

ProtFill beats the closest alternative on amino acid identity prediction by 13-25 percentage points and achieves comparable results in the recovery of atom coordinates.

here’s an example

redesigning differnet targets from PDB

ProtFill beats the closest alternative on amino acid identity prediction by 13-25 percentage points and achieves comparable results in the recovery of atom coordinates.

We generate two datasets using the proteinflow package: diverse (using all of PDB as the source) and antibody / nanobody (using SAbDab). For the experiments done on the antibody dataset, the masked regions are CDRs. The identity of the CDR is masked one at a time during training. For the diverse protein dataset we mask random areas on the complex interface.

Expanding beyond antibody applications

current model architectures are spoecifically designed for antibodies and cannot be used for general protein protein interactions

our model is general enough so that we cna mask any part of a protein structure

• in the case of antibodies, that’s the cdrs

• in the case of a different ppi, that’^s the interface

this gives us a better founddation to extend model capabilities towards solivng all kinds of protein protein interaction problems

we used protein flow again to generate a PPI test set and trained on that

ProtFill is also the first model that has been shown to work for redesigning binding interfaces up to 50 amino acids long for the general protein case.

Next works

Availability and code

ProtFill is freely available

Github: https://github.com/adaptyvbio/ProtFill