2021 CIKM |GF-VAE: A Flow-based Variational Autoencoder for Molecule Generation

Paper: https://dl.acm.org/doi/epdf/10.1145/3459637.3482260
Code: https://github.com/chshm/GF-VAE

GF-VAE: An automatic coder of molecular generation variation based on flow

Molecular generation is a challenging but meaningful work , He needs to follow the rules of chemical valence and optimize the given goal . Recently, the more effective method is the combination of molecular graph and generation model , But the cost of calculation is very high . therefore , The author puts forward GF-VAE, A stream based variational automatic encoder for molecular graph generation (VAE) Model . The model is in the original VAE That's an increase from Flow Model decoder . The encoder is mainly used to speed up the training of decoding , The decoder optimizes the performance of the encoder in turn . Because of the reversibility of the flow model , The generation process is easily accomplished by reversing the decoder . therefore ,GF-VAE Inherited VAE And the advantages of flow based methods . In molecular formation and reconstruction 、 Learn the smoothness of potential space 、 The model is verified by attribute optimization and constrained attribute optimization .

Model

In recent years , The generation model mainly includes generation of confrontation network (GAN)、 Variational automatic encoder (VAE) And flow based models . They have been applied to molecular generation , For example, based on GAN Of MolGAN and GCPN、 be based on VAE Of CVAE and JT-VAE, Flow based GraphNVP and MoFlow Model .

GF-VAE, yes VAE Seamless integration with normalized flow model , For one-time molecular graph generation .GF-VAE Use the flow generator to guide VAE Coders learn more meaningful advanced molecular representations , At the same time, it uses VAE The encoder reduces the weight of the stream generator . As shown in the figure below ：


$L_A/L_B$ Represents the entire module stack $L_A/L_B$ Time , For better mapping performance ,$K_A/K_B$ Represents the stacking times of the coupling layer in the corresponding box , The best combination of parameters is selected through subsequent experiments .

data

• QM9
  QM9 contain 134k molecular , most 9 Atoms , It is divided into 4 Different types .
• ZINC-250K
  ZINC-250K from 250k Molecular composition , most 38 Atoms , It is divided into 9 Different types .

The baseline

be based on VAE Model of

• CVAE
• GVAE
• GraphVAE

Flow based model

• GraphNVP
• GRF

Evaluation indicators

• Validity(V) The percentage of chemically effective molecules in all generated molecules ;
• Novelty(N) The percentage of effective molecules generated that do not appear in the training set ;
• Uniqueness(U) The only effective molecule accounts for the percentage of all generated molecules ;
• Reconstruction The percentage of molecules that can be reconstructed from their own potential vectors .
• Score(S), It is Validity、Novelty and Uniqueness The product of the

experiment

(a) and (b) Is from QM9 Two randomly sampled molecules . and (d) It shows the figure (a) and (b) Middle atom - Atomic cosine similarity heatmap, It is calculated according to the embedded vector of the encoder .(e) and (f) Shows based on GF-VAE The atomic embedding vector of the encoder calculates the similarity .

Qualitatively check the smoothness of the potential space of learning in two ways . One is to use the grid search method in two random orthogonal directions to find the neighborhood of randomly selected molecules in the potential space and visualize them . The other is to interpolate between two potential points of the molecular graph . choice Tanimoto Index as an index of chemical similarity , And pass heatmap Indicates their similarity value .

Upper figure （a） Show that the learned potential space is smooth , Because the adjacent potential points correspond to molecules with small changes .(b) It also shows that , In addition to the first step , The interpolation between the two potential points only slightly changes the molecular graph . This may be because the learned potential space does not follow a uniform distribution , Similar molecules are closely packed together , And different molecules are loosely dispersed around .

Molecular optimization

• One is feature optimization , It generates new molecules with the best characteristic fraction .
• The other is constrained property optimization , This means finding molecules that are similar to a given molecule but have better chemical properties .
  Choose a quantitative estimate of drug similarity (QED) and penalized logP(plogp) As a target attribute . Molecular similarity is measured by the Kurimoto similarity of Morgan fingerprints .

surface 4. stay ZINC-250K Top three scores plogP and QED Optimize molecules

surface 5.ZINC-250k Constraints on plogP Optimize

Constraint attribute optimization . The arrow points from the original molecule to the optimized molecule . The values on the left and right of the arrow represent the property improvement and similarity of a given molecular pair, respectively

Reference resources

https://baijiahao.baidu.com/s?id=1729293722854317823&wfr=spider&for=pc