Generating multiple and internal connections with LinkInvent

[Paul-Toth]

I've been messing around a lot with the LinkInvent model, and I was curious if there is an implementation for following use cases:

(1) Connecting Warheads with Multiple Attachment Points
*-C1=CC2=CC(-*)=CC=C2C=C1|C-1-*-C2=C(-*-1)C=CC=C2

(2) Building Internal Connections in a Single Molecule
*-C1=CC2=CC(CCCC3C-*-C4=C3C=CC=C4)=CC=C2C=C1

The reason I ask is that I am able to give these as the input warheads.smi file without causing any errors or warning in the log file. These specified molecules simply don't generate output. If I combine them with standard format warheads in the same input file, it successfully generates those normal compounds while not producing any output for my special cases.

I just want to make sure that there aren't additional options that might enable these input formats before I try to implement it myself.

For reference this is my input config:

run_type = "sampling"
device = "cuda:0"  # set torch device e.g. "cpu"
json_out_config = "_sampling.json"  # write this TOML to JSON

[parameters]

# Uncomment one of the comment blocks below.  Each generator needs a model
# file and possibly a SMILES file with seed structures.

## LinkInvent: find a linker/scaffold to link two fragments
model_file = "{reinvent_path}/priors/linkinvent.prior"
smiles_file = "{input_path}/warheads.smi"  # 2 warheads per line separated with '|'

output_file = 'samples.csv'  # sampled SMILES and NLL in CSV format

num_smiles = 200  # number of SMILES to be sampled, 1 per input SMILES
unique_molecules = true  # if true remove all duplicatesd canonicalize smiles
randomize_smiles = true # if true shuffle atoms in SMILES randomly

[halx]

Hi,

many thanks for your interest in REINVENT and welcome to the community!

It is not entirely clear what you are trying to achieve. If it is just about linking two molecules you need to have one * tag per molecule and delete all unneeded ones. We only support one linker in the LSTM model. If you want to create a cyclic system, you could try to run in two steps. Not sure how effective that would be but somebody here posted their own solution to generating macrocycles.

Cheers,
Hannes.

[Paul-Toth]

Thanks for the kind welcome!

And sorry if my descriptions were lacking, but you fully understood my goal!

I'll look into the macrocycle branch!

Best,
Paul T