readme.md

April 17, 2026 · View on GitHub

This repository contains code necessary to recapitulate the workflow described in this paper:

Connor V. Galvin, Amy B. Guo, Maple N. Chen, Isabella L. Alfonso, Dominic Grisingher, Simon Kretschmer, Divya Kranthi, Mark J. S. Kelly, Tanja Kortemme. A Hybrid Physics-Deep Learning Framework for Combinatorial De Novo Design of Small-Molecule Binding Proteins

where it was successfully applied to design binders for two target molecules, achieving high experimental success rates in both cases (4/26 tested designs bound target), sufficient to bypass the need for large-scale experimental screening.

Provided are scripts and instructions to perform motif library generation for an arbitrary small molecule target, along with matching, optimization, and evaluation of resultant designs. Excluded, however, is the generation of a protein scaffold library against which to screen binding motifs, for which the LUCS software was used in the referenced publication. The 1,816 LUCS-generated NTF2 fold proteins used therein is available as supplementary data, under "LUCS_NTF2_scaffold_library".

'workflow.py' provides a complete walkthrough of the workflow, with example terminal commands for executing every step as well as descriptions of inputs/outputs for each. Many of the scripts were produced according to the system specifications at the time of their application, and thus reference paths and distributed computing options that individual users must change to align with their system, but note is made where this is the case and the relevant lines are specified.

Necessary Software/Modules for 1:1 Execution of Workflow

prody
babel
openbabel
numpy
scipy
io
biopython
shutil
collections
pandas
amber antechamber
proteinMPNN
colabfold
Rosetta/pyrosetta

Overview of Workflow

Prepare target model/params
Define fragments/perform fragment search
Scrape pdb 4 contacts + filter
Cluster contacts
Generate motifs
RosettaMatch
Filter matches for ligand burial(+residue preferences)
EnzymeDesign application on match binding sites
HBond Refinement
Fast scoring - binding metrics (eg hbonds+burial)
Additional scoring binding metrics (preorganization, molecular contact surface, etc.)
ProteinMPNN on non binding site positions
Rosetta FastDesign with PMPNN sequence profile + score stability metrics
Colabfold
Score all metrics + final filtering of designs

Description of Contents

#db contains lists for matching user defined ligand substructures with existing ligands in pdb, then finding pdb id codes that those ligands are a part of to scrape contacts from

#workflow.py script includes information on general workflow with example code and flags used for submitting rosetta applications and xml scripts

#main contains the code for the motif library generation, ie scraping pdb for protein-ligand contacts with user defined ligand substructures, transforming relative to target ligand, filtering for quality+redundancy, and clustering the resultant interaction pools

#tools contains most useful design evaluation xmls and refinement scripts (hbrefine.py, fd_mpnn.xml)

#extra_tools contains additional scripts that may be useful, eg

-genpot folder contains params generation, enzyme design application + 	metric analysis, fastdesign (with and without 3bop), and coupledmoves 		application when using rosetta generic potential scorefunction 

-other_metric_eval contains rosetta xmls for design evaluation for different cases 

-cm_standard_jump1.xml = rosetta xml for running coupledmoves protocol 	with standard all atom scorefunction

-FD_XXXXX.xml scripts run rosetta fastdesign application with different options 

-motif_to_cst.py this script takes a pdb file as input and outputs a rosetta match cst file that describes the protein-ligand interction geometries in the input pdb 

-rfdaa_example.sh contains example command for running rosetta fold diffusion all atom and subsequently ligandmpnn on outputs