How to use property constraints for Mol2Mol sampling?

[enricogandini]

I am trying to understand how to use Mol2Mol that generate molecules that satisfy given property constraints.

By reading the Reinvent 4 paper, this should be possible. Table 4 of the paper lists Mol2Mol models originally described in paper Transformer-based molecular optimization beyond matched molecular pairs. These Mol2Mol variants should accept as input a property constraint (alongside the SMILES sequence), and generate new SMILES that satisfy the property constraint.

I read all the examples provided in this repository, and I searched the source code, but I don't understand how I can specify the property constraint in the configuration for a sampling run.

Please let me know if it's possible to use Mol2Mol in sampling mode while specifying property constraints. If I misunderstood the papers and this feature is not supposed to exist, please also let me know, I really want to understand how Mol2Mol works.

Thanks,
Enrico

[halx]

Hi,

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

The prior you are referring to is not available in REINVENT and I suggest to directly contact the authors of that paper. REINVENT uses reinforcement learning (RL) to restrain molecule generation to desired property. Multiple scoring components are readily available in the current release.

Many thanks,
Hannes.

[ankur-soley]

Hello,

I'm also working on optimizing a molecule to meet a set of property constraints. Could you please clarify which generator and TOML configuration format would be most suitable for this purpose?

[halx]

HI,

I think you would need to provide more detail as to how you want to optimize the molecule. Do you want to optimize around a scaffold or want to basically find nearest neighbours? The TOML configurations are independent of the generator.

Many thanks,
Hannes.

[ankur-soley]

Hi @halx,

Thank you for your reply.

I'm interested in generating optimized compounds around a given starting molecular structure, essentially focusing on molecular optimization around a scaffold, or ideally around the molecular topology (skeleton) to promote greater diversity in the generated compounds.

Based on the discussion in this thread #179, I'm currently using the staged_learning.toml template to generate new molecules. I would really appreciate your guidance or suggestions on the following points:

1. Choice of Prior
   I'm using the pubchem_ecfp4_with_count_with_rank_reinvent4_dict_voc.prior, as it appears to be the most comprehensive among the available priors for lead optimization tasks.

• Are there any other priors you would recommend for lead optimization?
• Could you also clarify how to select an appropriate value for the distance_threshold parameter when using the Mol2Mol prior?

2. PenalizeSameSmiles Behavior
   For molecule optimization tasks, PenalizeSameSmiles seems to be the most relevant. Could you confirm whether a lower penalty_multiplier value penalizes duplicate SMILES more or less?

Below is the toml file I'm currently using for testing:

run_type = "staged_learning"
device = "cuda:0"  # set torch device e.g. "cpu"
tb_logdir = "tb_logs"  # name of the TensorBoard logging directory
json_out_config = "_staged_learning.json"  # write this TOML to JSON

[parameters]

summary_csv_prefix = "staged_learning"  # prefix for the CSV file
use_checkpoint = false  # if true read diversity filter from agent_file
purge_memories = false  # if true purge all diversity filter memories after each stage

## Mol2Mol
prior_file = "/REINVENT4/priors/pubchem_ecfp4_with_count_with_rank_reinvent4_dict_voc.prior"
agent_file = "/REINVENT4/priors/pubchem_ecfp4_with_count_with_rank_reinvent4_dict_voc.prior"
smiles_file = "mol2mol.smi"  # 1 compound per line
sample_strategy = "multinomial"  # multinomial or beamsearch (deterministic)
distance_threshold = 100

batch_size = 64          # network

unique_sequences = true  # if true remove all duplicates raw sequences in each step
                         # only here for backward compatibility
randomize_smiles = true  # if true shuffle atoms in SMILES randomly

tb_isim = false  # track iSIM similarity in TensorBoard


[learning_strategy]

type = "dap"      # dap: only one supported
sigma = 128       # sigma of the RL reward function
rate = 0.0001     # for torch.optim


[diversity_filter]  # optional, comment section out or remove if unneeded
                    # NOTE: also memorizes all seen SMILES

type = "PenalizeSameSmiles" # IdenticalTopologicalScaffold,
                                 # ScaffoldSimilarity, PenalizeSameSmiles
bucket_size = 25                 # memory size in number of compounds
minscore = 0.4                   # only memorize if this threshold is exceeded
minsimilarity = 0.4              # minimum similarity for ScaffoldSimilarity
penalty_multiplier = 0.5         # penalty factor for PenalizeSameSmiles

### Stage 1
[[stage]]

chkpt_file = 'test1.chkpt'  # name of the checkpoint file, can be reused as agent

termination = "simple"  # termination criterion fot this stage
max_score = 1.0  # terminate if this total score is exceeded
min_steps = 25  # run for at least this number of steps
max_steps = 100  # terminate entire run when exceeded

[stage.scoring]
type = "arithmetic_mean"  # aggregation function

[[stage.scoring.component]]
# Custom alerts if used in a custom_product filter out unwanted groups
[stage.scoring.component.custom_alerts]

[[stage.scoring.component.custom_alerts.endpoint]]
name = "Unwanted SMARTS"  # user chosen name for output
weight = 0.79  # weight to fine-tune the relevance of this component

# parameters for the component:
# a list of unwanted SMARTS(!) to be scored as zero
params.smarts = [
    "[*;r8]",
    "[*;r9]",
    "[*;r10]",
    "[*;r11]",
    "[*;r12]",
    "[*;r13]",
    "[*;r14]",
    "[*;r15]",
    "[*;r16]",
    "[*;r17]",
    "[#8][#8]",
    "[#6;+]",
    "[#16][#16]",
    "[#7;!n][S;!$(S(=O)=O)]",
    "[#7;!n][#7;!n]",
    "C#C",
    "C(=[O,S])[O,S]",
    "[#7;!n][C;!$(C(=[O,N])[N,O])][#16;!s]",
    "[#7;!n][C;!$(C(=[O,N])[N,O])][#7;!n]",
    "[#7;!n][C;!$(C(=[O,N])[N,O])][#8;!o]",
    "[#8;!o][C;!$(C(=[O,N])[N,O])][#16;!s]",
    "[#8;!o][C;!$(C(=[O,N])[N,O])][#8;!o]",
    "[#16;!s][C;!$(C(=[O,N])[N,O])][#16;!s]"
]

[[stage.scoring.component]]
[stage.scoring.component.MolecularWeight]

[[stage.scoring.component.MolecularWeight.endpoint]]
name = "Molecular weight"  # user chosen name for output
weight = 0.342  # weight to fine-tune the relevance of this component

transform.type = "double_sigmoid"
transform.high = 450.0
transform.low = 200.0
transform.coef_div = 450.0
transform.coef_si = 20.0
transform.coef_se = 20.0

[halx]

Yes, the prior and diversity filter you have chosen are appropriated. The distance_threshold is a purely empirical parameter and there is no right or wrong value for it.