how to generate the new/novel molecules by using TL_Mol2Mol.model after running transfer learning?

[MachineGUN001]

Hi,

sorry to bring a naive question.

after running the transfer learning by command line $ reinvent -l tl.log transfer_learning.toml, the train.csv and output model named as TL_Mol2Mol.model via toml setting output_model_file = "./models/TL_Mol2Mol.model", were generated.

how can I generate the new/novel molecules by using TL_Mol2Mol.model?

How should I set the parameters of the config file (e.g. toml) if I want to get new compounds with more diversity of scaffold variations?

could you please provide a demo or suggestions ?
I know your time is valuable, but your help would be greatly appreciated.

many many thanks

Sh-Y

[halx]

Hi again,

to generate molecules from a model you would need to sample from it. There is a special "sampling" runmode for this with example config files in the repository,

However, that being said, the problem with TL is that it is hard to control: if you run for too short the model will stay close to the original prior, if you run for too long the generated molecules may be very close to the input data set. Neither is desirable in practice. TL is therefore best be used to "focus" a prior model closer to the desirect structures you expect to receive. But then you would use RL with a clearly defined "scoring profile" to optimizxe the model for this profile. In practice, 3D methods like docking or shape similarity give better results than just 2D or simple physchem properties.

Hope this helps,
Hannes.

[MachineGUN001]

hi, Hannes,

Thank you very much for your help, and I have know more about the work. Can you provide a separate demonstration toml for the xxx you mentioned, or HIGHLIGHT out the necessary parameters? thank you very much!

this is my toml for sampling, Can you help me see if this is the right way to write it? If I want to be able to generate more diverse molecules, how should I set the parameters?

run_type = "sampling"
use_cuda = true  # run on the GPU if true, on the CPU if false
json_out_config = "_sampling.json"  # write this TOML to JSON


[parameters]
## Mol2Mol: find molecules similar to the provided molecules
model_file = "./models/TL_Mol2Mol.model" # my trained model after transfer learning
smiles_file = "./data/test.smi"  # 1 compound per line
sample_strategy = "beamsearch"  # multinomial or beamsearch (deterministic)
temperature = 1.0 # temperature in multinomial sampling
tb_logdir = "tb_logs"  # name of the TensorBoard logging directory

output_file = 'sampling001.csv'  # sampled SMILES and NLL in CSV format

num_smiles = 157  # 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

many many thanks

Best,
Sh-Y

[halx]

The config file looks fine to me. It may not be best to use the final model from TL as it could be too much biased towards your input molecules. As said it is hard to judge when TL should be stopped (TL writes "intermediate" model files at defined steps so you can pick up from there). The question at the end is what you end goal is.