Source-linked AI summary
Application of generative autoencoder in de novo molecular design
Thomas Blaschke, Marcus Olivecrona, Ola Engkvist, Jürgen Bajorath, Hongming Chen
TL;DR
Generating novel molecules with desirable pharmacological and physicochemical properties remains a challenge. The paper evaluates generative autoencoders that map molecules to continuous latent spaces and back, then searches those spaces for target-guided structures. The latent spaces preserve chemical similarity and support generation of compounds predicted active against DRD2, including compounds similar to known actives outside training.
Problem
Inverse-QSAR design requires mapping predicted molecular descriptors back to chemical structures, a difficult task for nonlinear models.
Method
The study compares four autoencoder architectures and uses Bayesian optimization to search their latent spaces with QSAR-guided target functions.
Results
The latent space preserves chemical similarity, and systematic search generated novel compounds predicted active against DRD2 and similar to validated actives.
Takeaways & Limitations
Autoencoder latent spaces can be used to generate analogue structures and guide de novo molecular design toward compounds predicted active by a QSAR model.
Takeaways & Limitations
The similarity-based search depends on preserving the similarity principle, and target-guided optimization generated large macrocycles that generally have low synthetic feasibility.
Abstract
from arXiv · showhide
A major challenge in computational chemistry is the generation of novel molecular structures with desirable pharmacological and physiochemical properties. In this work, we investigate the potential use of autoencoder, a deep learning methodology, for de novo molecular design. Various generative autoencoders were used to map molecule structures into a continuous latent space and vice versa and their performance as structure generator was assessed. Our results show that the latent space preserves chemical similarity principle and thus can be used for the generation of analogue structures. Furthermore, the latent space created by autoencoders were searched systematically to generate novel compounds with predicted activity against dopamine receptor type 2 and compounds similar to known active compounds not included in the training set were identified.
1 Introduction
Deep learning generative models offer a route to novel molecular structures, addressing inverse-QSAR challenges by generating compounds directly from desired properties. This study evaluates autoencoders for latent-space molecular generation and target-guided design.
- 1 Introduction: Deep learning has been applied to drug discovery, including QSAR modeling and generative models that produce novel chemical structures.RNN-based models learn SMILES character distributions, while QSAR models predict molecular properties.
- 1 Introduction: Inverse-QSAR methods face a major challenge because descriptor values must be explicitly mapped back to chemical structures.Existing approaches are limited for nonlinear QSAR models or make novel-compound identification difficult through virtual screening.
- 1 Introduction: Variational autoencoders map compounds into a latent space that can be sampled or navigated to generate structures with desired chemical properties.Prior work showed latent-space generation, although targeted searching was difficult in one organic light-emitting diode case study.
- 1 Introduction: This study compares multiple autoencoder models for structure generation and uses Bayesian optimization to search latent space for compounds predicted active by a QSAR model.The strategy was applied to generate structures predicted to be active against a defined target.
2.1 Neural networks
Neural networks learn representations or transformations of data through interconnected layers, with architectures tailored to reconstruction, sequence processing, or pattern extraction. Autoencoders compress inputs into latent representations and reconstruct them, while recurrent and convolutional networks process sequential and local patterns.
- 2.1 Neural networks: An autoencoder uses an encoder, decoder, and distance function to compress high-dimensional input into a latent space and minimize reconstruction information loss.Because reconstruction targets come from the inputs, the model is regarded as self-supervised.
- 2.1 Neural networks: The latent-space bottleneck forces an autoencoder to extract informative features, with performance depending on the data and encoder-decoder architecture.The basic autoencoder is sometimes described as nonlinear PCA.
- 2.1 Neural networks: A variational autoencoder represents the continuous code as a probabilistic latent variable, using probabilistic encoding and decoding distributions.The formulation imposes a prior distribution on the latent representation.
- 2.1 Neural networks: A recurrent neural network applies the same function across sequential elements while using previous outputs to influence current and future states.LSTM and GRU architectures regulate which prior information is retained or discarded.
- 2.1 Neural networks: A convolutional neural network extracts patterns by applying shared convolution operations that combine neighboring input data.CNNs include convolutional, pooling, and fully connected hidden layers.
2.2 Implementation details
The implementation uses autoencoders to map molecular SMILES into latent representations and decode latent points into valid molecular structures. VAE and AAE variants differ in how they regularize the latent distribution, while generation samples decoder token probabilities and validates resulting SMILES.
- Variational autoencoder: The basic autoencoder maps molecules into continuous latent space, but arbitrary latent points may not decode because the model can memorize training examples.The VAE addresses this by treating the representation as a probabilistic latent variable with an imposed prior distribution.
- Variational autoencoder: VAE training maximizes reconstruction likelihood while regularizing qφ(z|X) toward the prior p(z) through Kullback–Leibler divergence minimization.The prior is described as a multivariate Gaussian with zero mean and unit variance in the loss formulation.
- Variational autoencoder: The VAE encoder uses CNN and fully connected layers, while its decoder uses a fully connected layer followed by three GRU-based RNN layers.The final GRU layer produces token probabilities; teacher forcing supplies target tokens during training, whereas generation feeds sampled previous tokens.
- Adversarial autoencoder: The AAE adds a discriminator so the encoder output follows a chosen target distribution while the decoder minimizes reconstruction error.The discriminator distinguishes encoder-generated representations from samples drawn from the target prior, and the encoder is trained to fool it.
- Adversarial autoencoder: AAE training iteratively minimizes reconstruction loss, trains the discriminator to distinguish prior samples from encoder outputs, and trains the encoder to fool the discriminator.These three steps are repeated for each batch until the reconstruction loss converges; teacher forcing is always used for AAE decoders.
- Tokenizing SMILES and generating new SMILES: SMILES are tokenized into 35 tokens, padded to 120 tokens, and converted to one-hot inputs before autoencoder training.Structures are canonicalized with RDKit, stereochemistry is removed, and molecules with fewer than 10 heavy atoms or more than 120 tokens are excluded.
- Tokenizing SMILES and generating new SMILES: For each latent point, the decoder samples 500 sequences, retains valid SMILES checked with RDKit, and assigns the most frequent valid sequence as output.Bayesian optimization searches latent space by maximizing a target-activity score predicted by Gaussian-process models trained from 100 starting points.
3 Results and Discussion
The autoencoders were compared as molecular structure generators, then assessed for similarity preservation and target-activity-guided search in latent space. The Uniform AAE produced a smooth latent representation that supported searches yielding compounds predicted active against DRD2.
- Model comparison: Four autoencoder types were trained and compared: NoTeacher VAE, Teacher VAE, Gauss AAE, and Uniform AAE.The models differed in teacher forcing and the encoder's target latent distribution.
- Structure generation: Teacher-forcing models generated a much higher fraction of valid SMILES than NoTeacher VAE despite its higher character reconstruction accuracy.NoTeacher VAE errors often disrupted matching branching or ring symbols, producing invalid SMILES.
- Similarity preservation: Generated structures generally became less chemically similar to Celecoxib as their latent-space distance increased across all four latent spaces.This relationship supported similarity searches around a query molecule.
- Similarity preservation: After Celecoxib analogues with FCFP4 Tanimoto similarity above 0.5 were removed from training, all models still reconstructed Celecoxib and close analogues at distances below 1.This tested similarity preservation when the query and close analogues were excluded from training.
- Similarity preservation: At Euclidean distance 4, Uniform AAE generated more than 30% valid SMILES, while at distance 6 it generated 20% versus 5% for Teacher VAE.The authors interpret this as evidence that Uniform AAE provides the smoothest latent chemical-space representation.
- Target-activity-guided generation: 369591 compounds sampled from Bayesian-optimization solutions with average Pactive above 0.95 included 11.5% with ECFP6 Tanimoto similarity above 0.35 to the nearest validated active.The generated compounds were predicted highly active, with Pactive > 0.99, and mostly shared chemical scaffolds with validated actives; structures with rings larger than eight atoms were filtered.
4 Conclusion
The study applies generative adversarial autoencoders to inverse QSAR, using latent-space search to identify novel structures predicted active against DRD2. The generated latent space preserves local chemical similarity and supports analogue generation.
- The study introduces generative adversarial autoencoder neural networks for inverse QSAR and novel chemical-structure generation.
- Autoencoder latent spaces preserve local similarity and support analogue generation around query structures such as Celecoxib.
- Bayesian optimization searches the latent space under QSAR guidance to identify novel structures predicted active against DRD2.
Funding
The study reports funding from the European Union’s Horizon 2020 program and discloses employment affiliations for several authors.
- Thomas Blaschke received Horizon 2020 funding through the Marie Sklodowska-Curie grant agreement No 676434.
- The funding acknowledgment identifies the Big Data in Chemistry project, BIGCHEM.
- Marcus Olivecrona, Hongming Chen, and Ola Engkvist are employed by AstraZeneca, while Jürgen Bajorath is employed by the University of Bonn.