MoLeR 分子生成模型，可以指定包含骨架scaffold生成

参考：
https://github.com/microsoft/molecule-generation
https://arxiv.org/pdf/2103.03864.pdf

安装：
注意需要安装tf依赖

pip install molecule-generation

1、使用代码

权重下载，百度网盘 https://pan.baidu.com/s/1lkiWK9-d5MvNyzqRrusGXA?pwd=4hij；保存到example_model_directory文件下

from molecule_generation import load_model_from_directory

model_dir = "./example_model_directory"
example_smiles = ["c1ccccc1", "CNC=O"]

with load_model_from_directory(model_dir) as model:
    embeddings = model.encode(example_smiles)
    print(f"Embedding shape: {embeddings[0].shape}")

    # Decode without a scaffold constraint.
    decoded = model.decode(embeddings)

    # The i-th scaffold will be used when decoding the i-th latent vector.
    decoded_scaffolds = model.decode(embeddings, scaffolds=["CN", "CCC"])

    print(f"Encoded: {example_smiles}")
    print(f"Decoded: {decoded}")
    print(f"Decoded with scaffolds: {decoded_scaffolds}")

embeddings 产生的是512维度；

没指定骨架生成

with load_model_from_directory(model_dir) as model:
    embedding2 = model.encode(["O=C1[C@@H](CC)N(CC2=CC(C)=CS2)C3=C(N1C)C=NC(NC4=C(OCC)C=C(C5CCN(C)CC5)C=C4)=N3"])

    
    # The i-th scaffold will be used when decoding the i-th latent vector.
    decoded2 = model.decode(embedding2, scaffolds=["N1C(=O)c2ccc(Br)cc2C1=O"])
    print(decoded2)

指定骨架生成

with load_model_from_directory(model_dir) as model:
    embedding2 = model.encode(["O=C1[C@@H](CC)N(CC2=CC(C)=CS2)C3=C(N1C)C=NC(NC4=C(OCC)C=C(C5CCN(C)CC5)C=C4)=N3"])

    
    # The i-th scaffold will be used when decoding the i-th latent vector.
    decoded2 = model.decode(embedding2, scaffolds=["N1C(=O)c2ccc(Br)cc2C1=O"])
    print(decoded2)

生成多个相似分子，latent采样附近来生成

1）指定骨架

import numpy as np
from molecule_generation import VaeWrapper

model_dir = "./example_model_directory"
scaffold = "C1=CC=CC=C1"
init_mol = "O=CNC1=CC=CC=C1"

simless =[]

with VaeWrapper(model_dir) as model:
    [latent_center] = model.encode([init_mol])
    latents = latent_center + 0.5 * np.random.randn(10, latent_center.shape[0]).astype(np.float32)  ### 一个latent_center中心，随机latent采样附近10个来生成
    # print("latents:",latents)

    for idx, smiles in enumerate(model.decode(latents, scaffolds=[scaffold] * len(latents))):
        print(f"Result #{idx + 1}: {smiles}")
        simless.append(smiles)

from rdkit.Chem import AllChem, Draw

Draw.MolsToGridImage([ Chem.MolFromSmiles(i) for i in simless],subImgSize=(300,300), molsPerRow=5)  

2）不指定骨架

import numpy as np
from molecule_generation import VaeWrapper

model_dir = "./example_model_directory"
init_mol = "O=C1[C@@H](CC)N(CC2=CC(C)=CS2)C3=C(N1C)C=NC(NC4=C(OCC)C=C(C5CCN(C)CC5)C=C4)=N3"

simless =[]

with VaeWrapper(model_dir) as model:
    [latent_center] = model.encode([init_mol])
    print("latent_center:",latent_center)
    latents = latent_center + 0.5 * np.random.randn(10, latent_center.shape[0]).astype(np.float32)
    print("latents:",latents)
    print([scaffold] * len(latents))
    for idx, smiles in enumerate(model.decode(latents)):
        print(f"Result #{idx + 1}: {smiles}")
        simless.append(smiles)