AI-assisted protein design to rapidly convert antibody sequences to intrabodies targeting diverse peptides and histone modifications 

Set up:

This was written with cuda 11.8 in mind, but should easily be exportable to any other cuda version; no custom GPU development was done. The following major packages are dependencies:

• ANARCI (https://github.com/oxpig/ANARCI)
• localcolabfold (YoshitakaMo/localcolabfold: ColabFold on your local PC)
• ProteinMPNN (dauparas/ProteinMPNN: Code for the ProteinMPNN paper)

And there are the following python dependencies:

• Python 3.9
• Biopython (should get installed when you install ANARCI)
• NumPy (should get installed with both ANARCI and ProteinMPNN)
• Matplotlib
• Scipy
• PyTorch (should get installed with ProteinMPNN)
• Web Logo

A requirements.txt file is not provided because 99% of the setup for this project is getting ANARCI, localcolabfold, and ProteinMPNN set up and operational in a singular environment. The extraneous libraries are very easy to install after getting the major 3 operational. If you find it useful, the conda env for this project is provided in environment.yaml - but again, it is highly recommended to install each of the 3 major dependicies independently, and not with this .yaml file.

Usage

This pipeline was written with ease of use in mind. There are two scripts worth mentioning here. The first being the scFv-ification of antibody sequences (scripts/scfv_anarci.py), and the second being the actual run pipeline itself (af2_pmpnn.sh).

scFv-ification of an antibody

An scFv (single chain variable fragment) is an antibody with only the CDR loops and variable region framework, of both the constant and heavy chains. A flexible linker is then used to combine the remaining regions to create and a single chain domain. The process of doing this is tedious and time consuming by hand; identify the CDR regions, identify the variable and constant regions, and stitch them together. For both speed and consistency, a script was written to quickly automate this process.

Let’s use the hemagglutinin (HA) Heavy and Light chains for our example, called HA.fasta :

>HA_heavy
MKLPVLLVVLLLFTSPASSSEVKLVESGGDLVKPGGSLKLSCAASGFTFSSYGMSWVRQTPDKRLEWVATISRGGSYTYYPDSVKGRFTISRDNAKNTLYLQMSSLKSEDTAMYYCARRETYDEKGFAYWGQGTTVTVSSARPTAPSVYPLAPVCGDTTGSSVTLGCLVKGYFPEPVTLTWNSGSLSSGVHTFPAVLQSDLYTLSSSVTVTSSTWPSQSITCNVAHPASSTKVDKKIEPRGPTIKPCPPCKCPAPNLLGGPSVFIFPPKIKDVLMISLSPIVTCVVVDVSEDDPDVQISWFVNNVEVHTAQTQTHREDYNSTLRVVSALPIQHQDWMSGKEFKCKVNNKDLPAPIERTISKPKGSVRAPQVYVLPPPEEEMTKKQVTLTCMVTDFMPEDIYVEWTNNGKTELNYKNTEPVLDSDGSYFMYSKLRVEKKNWVERNSYSCSVVHEGLHNHHTTKSFSRTPGK
>HA_light
MTSTLPFSPQVSTPRSKFATMEFQTQVLMSLLLCMSGAAADIELTQSPSSLTVTAGEKVTMSCKSSQSLLNSGNQKNYLTWYQQKPGQPPKLLIYWASTRESGVPDRFTGSGSGRDFTLTISSVQAEDLAVYYCQNDNSHPLTFGAGTKLELKRADAAPTVSIFPPSSEQLTSGGASVVCFLNNFYPKDINVKWKIDGSERQNGVLNSWTDQDSKDSTYSMSSTLTLTKDEYERHNSYTCEATHKTSTSPIVKSFNRNEC

Identifing the variable regions, and stitching them together with the default flexible linker (GGGGS x3): scfv_anarci.py HA.fasta Yeilds the scFv sequence printed to the screen: MEVKLVESGGDLVKPGGSLKLSCAASGFTFSSYGMSWVRQTPDKRLEWVATISRGGSYTYYPDSVKGRFTISRDNAKNTLYLQMSSLKSEDTAMYYCARRETYDEKGFAYWGQGTTVTVSSGGGGSGGGGSGGGGSDIELTQSPSSLTVTAGEKVTMSCKSSQSLLNSGNQKNYLTWYQQKPGQPPKLLIYWASTRESGVPDRFTGSGSGRDFTLTISSVQAEDLAVYYCQNDNSHPLTFGAGTKLELK

If you’d like to use any other sequence besides the default, this is easily changable with the —linker-seq flag. Note that this DOES NOT check for legitimate amino acids:

scfv_anarci.py HA.fasta --linker-seq GSGSGSGSGSGSGSG MEVKLVESGGDLVKPGGSLKLSCAASGFTFSSYGMSWVRQTPDKRLEWVATISRGGSYTYYPDSVKGRFTISRDNAKNTLYLQMSSLKSEDTAMYYCARRETYDEKGFAYWGQGTTVTVSSGSGSGSGSGSGSGSGDIELTQSPSSLTVTAGEKVTMSCKSSQSLLNSGNQKNYLTWYQQKPGQPPKLLIYWASTRESGVPDRFTGSGSGRDFTLTISSVQAEDLAVYYCQNDNSHPLTFGAGTKLELK

scfv_anarci.py HA.fasta --linker-seq __science_is_cool__ MEVKLVESGGDLVKPGGSLKLSCAASGFTFSSYGMSWVRQTPDKRLEWVATISRGGSYTYYPDSVKGRFTISRDNAKNTLYLQMSSLKSEDTAMYYCARRETYDEKGFAYWGQGTTVTVSS__science_is_cool__DIELTQSPSSLTVTAGEKVTMSCKSSQSLLNSGNQKNYLTWYQQKPGQPPKLLIYWASTRESGVPDRFTGSGSGRDFTLTISSVQAEDLAVYYCQNDNSHPLTFGAGTKLELK

If you’d like to save your sequence to a file, use the --output flag. scfv_anarci.py HA.fasta --output HA_scfv Will produce the file HA_scfv.fasta with sequence title >HA_scfv. And if we look inside that file:

$ cat HA_scfv.fasta

>HA_scfv
MEVKLVESGGDLVKPGGSLKLSCAASGFTFSSYGMSWVRQTPDKRLEWVATISRGGSYTYYPDSVKGRFTISRDNAKNTLYLQMSSLKSEDTAMYYCARRETYDEKGFAYWGQGTTVTVSSGGGGSGGGGSGGGGSDIELTQSPSSLTVTAGEKVTMSCKSSQSLLNSGNQKNYLTWYQQKPGQPPKLLIYWASTRESGVPDRFTGSGSGRDFTLTISSVQAEDLAVYYCQNDNSHPLTFGAGTKLELK

scFv Framework Optimization

Now that we’ve successfully created an scFv sequence, we can optimize its framework with the pipeline that is the crux of the protein design aspect of the paper. The heart and soul of the pipeline is the af2_pmpnn.sh bash script. In it are a handful of input variables that we’ll go over briefly.

folder_with_pdbs="input_dir" Is the input directory where all pdbs or fastas (one scFv per fasta!) we want to redesign the framework of should go.

seqs_per_run=3 is the number of sequences to generate via ProteinMPNN for the scFv framework.

determine_CDRs="martin" Is the method with which to determine the CDRs of the scFv. There are four supported options here: martin, kabat, and chothia are all self explanatory. These are the sequence numbering schemes with which the literature has widly used to identify CDRs in an antibody. The fourth and final one is structure . This takes a structure based approach (as opposed to sequence for the other 3) to identifying the CDR loops. A peptide and the scFv have their structure predicted via AF2, then all the residues within 6 angstroms of that peptide are considered the “CDRs”, and they are not allowed to mutate.

simple_grab=true is used to determine how do we determine what residues are near the CDRs. For example, if two residues are near a CDR and create a “bridge” if 1 residue that isn’t, this residue would be allowed to change if simple_grab is set to true. However, if we set this to false, then that residue would artificially be considered part of the group of residues near the CDR loops that are not allowed to change.

output_dir="output_dir” the name of the output directory

to_design="framework" What are we redesigning during this run? This was put in to help with other projects from the Snow, Geiss, and Stasevich labs. to_design is only allowed to be 1 of 3 things: framework to redesign the framework of the scFv, loops to redesign the loops of the scFv, and lss to redesign both the loops and the secondary shell of the loops (all residues within 4A of the loops, modified by simple_grab).

PMPNN='path/to/my/ProteinMPNN' Path to the ProteinMPNN directory

In the output directory, there are 4 key files to look at (continuing our HA example): HA_noWT.fa contains all of the framework designed sequences that are produced. HA_sequence_log.png is an image of the sequence logo of the designed framework. HA_plddt.png is a plot of the pLDDT’s (pLDDT: Understanding local confidence | AlphaFold) for each scFv. HA_ptm.png is a plot of the pTM’s (Confidence scores in AlphaFold-Multimer | AlphaFold) for each scFv.

To make our sequence selections, we took the top sequences that performed well in both pLDDT and pTM.

TODO

Dockerfile / Docker image