Source-linked AI summary
Molecular generative model based on conditional variational autoencoder for de novo molecular design
Jaechang Lim, Seongok Ryu, Jin Woo Kim, Woo Youn Kim
TL;DR
Molecular design must search an enormous chemical space, while conventional screening and structural optimization may not control correlated properties simultaneously. The paper introduces a conditional variational autoencoder that conditions molecular generation on target properties. It generated drug-like molecules satisfying five properties, enabled selective property adjustment, and extended properties beyond the training range, although success rates were low and SMILES omitted 3D conformational information.
Problem
Molecular design spans an estimated 10^23 to 10^60 drug-like molecules, making experiments alone insufficient and motivating computational generation beyond conventional screening.
Method
The CVAE incorporates molecular properties as a condition vector in both encoding and decoding to control multiple properties during molecule generation.
Results
The model generated drug-like molecules satisfying five target properties within 10% error, selectively adjusted LogP without changing other properties, and generated LogP and TPSA beyond the training range.
Takeaways & Limitations
The demonstrated CVAE strategy directly produces molecules with desirable properties and offers simultaneous and selective property control for de novo molecular design.
Takeaways & Limitations
The success rate for generating desirable molecules was very low, and SMILES lacks 3D conformational information needed when conformational effects are critical.
Abstract
from arXiv · showhide
We propose a molecular generative model based on the conditional variational autoencoder for de novo molecular design. It is specialized to control multiple molecular properties simultaneously by imposing them on a latent space. As a proof of concept, we demonstrate that it can be used to generate drug-like molecules with five target properties. We were also able to adjust a single property without changing the others and to manipulate it beyond the range of the dataset.
Introduction
Molecular design is difficult because the drug-like chemical space is vast and experiments alone cannot efficiently identify molecules with desired properties. The paper proposes a conditional variational autoencoder to control multiple correlated molecular properties during generation.
- Introduction: Only 10^8 molecules have been synthesized, compared with an estimated 10^23 to 10^60 drug-like molecules.
- Introduction: Virtual screening reduces experimental effort, but generated or library molecules may still fail to meet the required criteria.
- Introduction: Latent-space generative models make gradient-based molecular-property optimization feasible by representing molecules as real vectors.
- Introduction: Because molecular properties are correlated, modifying one property can undesirably change others, motivating simultaneous multivariable control.
- Introduction: The CVAE generated drug-like molecules meeting five target properties, adjusted one property while preserving others, and generated properties beyond the database range.
Conditional variational autoencoder (CVAE)
The CVAE extends the VAE by conditioning both encoding and decoding on desired molecular properties. This design embeds target properties directly in the latent representation, reducing the need for separate property optimization.
- Conditional variational autoencoder (CVAE): The VAE combines reconstruction error and KL divergence terms while encoding data into, and decoding molecules from, a latent space.
- Conditional variational autoencoder (CVAE): The CVAE adds conditional information to the VAE objective and directly incorporates a condition vector into encoding and decoding.
- Conditional variational autoencoder (CVAE): In this model, controlled molecular properties form the condition vector, allowing target properties to be imposed during molecule generation.
- Conditional variational autoencoder (CVAE): Unlike the jointly trained VAE approach, the CVAE directly incorporates properties into both encoder and decoder and does not require further optimization for each target.
- Conditional variational autoencoder (CVAE): The latent vector separates target-property information from molecular-structure and other-property information, except where properties are strongly coupled to a scaffold.
Molecular representation and model construction
Molecules are represented as canonicalized SMILES sequences and molecular properties are encoded in a condition vector. An LSTM-based recurrent architecture processes these representations through the CVAE.
- Molecular representation and model construction: Canonicalized SMILES codes represent molecules as unique sequences, with an end marker and one-hot vectors for each character.
- Molecular representation and model construction: The condition vector stores MW, LogP, and TPSA as normalized values, while HBD and HBA use one-hot representations.
- Molecular representation and model construction: A three-layer RNN with LSTM cells and 500 hidden nodes per layer is used for both the encoder and decoder.
- Molecular representation and model construction: The decoder uses softmax outputs and cross entropy for the reconstruction-error cost.
Dataset and hyperparameters
The study uses 500,000 ZINC molecules to evaluate a CVAE that controls five molecular properties and generates molecules within and beyond the dataset’s property range.
- Dataset: 500,000 molecules from ZINC formed the dataset, with 80% used for training.Using 5,000,000 molecules did not improve validation or target-generation success rates compared with 500,000 molecules.
- Targeted generation: Nine molecules generated for each of Aspirin and Tamiflu matched their five target properties within a 10% error range.The generated structures differed considerably from the original molecules because latent vectors were sampled with noise.
- Structure-preserving generation: Latent-vector sampling around Aspirin produced molecules similar in both structure and properties within a 10% error range.This application used latent vectors sampled around the mother molecule rather than random training-set vectors.
- Single-property control: LogP varied from -0.23 to 3.55 while the other four Tamiflu properties remained similar, although LogP beyond 4.0 was not generated.Independent adjustment of MW and TPSA was difficult because those properties are highly correlated.
- Beyond-range generation: The model generated molecules with LogP larger than 5.5 and TPSA larger than 165.5, but validity was relatively low outside the training range.The generated-property distributions shifted toward larger values compared with randomly selected training molecules.
- Model inputs: The CVAE controlled molecular weight, LogP, HBD, HBA, and TPSA through a condition vector.These properties were used to generate drug-like molecules matching specified targets.