Source-linked AI summary
Conditional molecular design with deep generative models
Seokho Kang, Kyunghyun Cho
TL;DR
Efficiently exploring chemical space while generating molecules with desired properties remains challenging. The paper presents a conditional molecular-generation approach using SSVAE and reports effective property prediction and molecule generation without extra optimization.
Problem
Efficiently exploring a large chemical space while proposing molecules with desired properties remains challenging.
Method
The approach uses an SSVAE model for conditional molecular generation and property prediction.
Results
The model demonstrates effectiveness in property prediction and conditional molecular design, efficiently generating novel molecules satisfying target conditions without extra optimization.
Takeaways & Limitations
The approach can serve as an efficient tool for designing new chemical structures.
Takeaways & Limitations
The application is limited to the portion of chemical space representable by SMILES.
Abstract
from arXiv · showhide
Although machine learning has been successfully used to propose novel molecules that satisfy desired properties, it is still challenging to explore a large chemical space efficiently. In this paper, we present a conditional molecular design method that facilitates generating new molecules with desired properties. The proposed model, which simultaneously performs both property prediction and molecule generation, is built as a semi-supervised variational autoencoder trained on a set of existing molecules with only a partial annotation. We generate new molecules with desired properties by sampling from the generative distribution estimated by the model. We demonstrate the effectiveness of the proposed model by evaluating it on drug-like molecules. The model improves the performance of property prediction by exploiting unlabeled molecules, and efficiently generates novel molecules fulfilling various target conditions.
Introduction
Molecular design seeks novel molecules with desired properties, but efficiently exploring large chemical spaces is difficult. Existing prediction-and-generation workflows rely on screened candidates or additional latent-space optimization, motivating direct conditional generation.
- Introduction: Property-prediction workflows screen manually obtained candidate molecules, requiring sufficient labeled data because accuracy depends on label quantity and quality.Candidates may come from fragment enumerations and public databases.
- Introduction: Generation models learn a molecular latent space and randomly produce molecules analogous to the training set for subsequent property screening.Prior systems commonly use SMILES with recurrent neural networks, VAEs, or GANs with recurrent decoders.
- Introduction: Conditional molecular design targets molecules whose properties are close to predetermined conditions by finding latent representations reflecting those conditions.Previous approaches include recursive fine tuning, Bayesian optimization, and reinforcement learning.
- Introduction: Previous conditional methods generate intended properties indirectly through additional optimization, which is inefficient when multiple target conditions are considered.The optimization is often performed in latent space.
- Introduction: The proposed model combines property prediction and molecule generation in a semi-supervised variational autoencoder for direct conditional molecular design.New molecules are sampled directly from a conditional generative distribution without extra optimization, while unlabeled molecules are exploited when only a small portion is labeled.
Methods
The method uses a semi-supervised variational autoencoder to jointly predict molecular properties and generate SMILES molecules, learning from labeled and unlabeled data. Conditional generation fixes target properties, samples remaining property values and latent variables, and decodes molecules with beam search.
- Model architecture: The model adapts a semi-supervised variational autoencoder for continuous molecular-property outputs, treating missing properties as latent variables predicted by qφ(y|x).The framework models molecules as x and their continuous-valued properties as y.
- Input representation: Molecules are represented as SMILES sequences of one-hot symbol vectors, while each property vector contains m scalar values.Two-character atoms such as Si, Cl, Br, and Sn are treated as single symbols in the vocabulary.
- Model architecture: Three RNNs implement property prediction qφ(y|x), encoding qφ(z|x,y), and decoding pθ(x|y,z), with bidirectional predictor and encoder networks.The decoder generates sequences autoregressively from previous outputs, y, and z.
- Training objective: Training combines variational objectives for labeled and unlabeled molecules with a supervised mean squared error term weighted by β.β controls the trade-off between generative learning and supervised learning; β = 0 gives fully generative learning, while larger β emphasizes supervised learning.
- Molecule generation: Beam search approximates optimal autoregressive decoding because exact search is computationally intractable and independent sampling is vulnerable to sequence-generation noise.Beam search maintains top-K candidates at each time step while generating sequences from left to right.
Results and Discussion
The study evaluates SSVAE on 310,000 drug-like molecules using partially labeled property data. SSVAE improves property prediction with fewer labels and generates molecules whose properties fulfill specified target conditions.
- Dataset: The dataset contains 310,000 drug-like SMILES molecules, with 300,000 used for training and 10,000 for testing.The evaluated properties are molecular weight (MolWt), LogP, and QED.
- Model and evaluation: Property prediction is evaluated by MAE while varying the labeled fraction across 5%, 10%, 20%, and 50% of the training set.Baseline comparisons include ECFP, GraphConv, an independently trained predictor, and VAEproperty.
- Model and evaluation: SSVAE is trained jointly on labeled and unlabeled molecules, combining property prediction with variational molecule generation.The model uses predictor, encoder, and decoder networks and minimizes the objective in Equation 3.
- Property prediction: SSVAE outperformed baseline models in most cases and achieved better prediction performance than the predictor model with fewer labeled molecules.The performance difference narrowed as the labeled fraction increased.
- Conditional molecular design: SSVAE generated new molecules fulfilling target conditions, with conditional property distributions centered near target values and smaller standard deviations.Conditional outputs followed training-set property distributions for molecules near the target value.
- Conditional molecular design: The conditional-design accuracy tended to track property-prediction accuracy: MolWt distributions were narrow and distinct, whereas LogP and QED showed more overlap.The normalized MAEs for MolWt, LogP, and QED were 0.016, 0.038, and 0.058, respectively.
Conclusion
The SSVAE conditionally generates novel molecules from a partially labeled training set and supports property prediction and design without extra optimization. Its applicability remains bounded by SMILES coverage, and generated structures require further examination for realism.
- Conclusion: The SSVAE was trained on a partially labeled molecular dataset, with properties available for only a small portion of molecules.New molecules are sampled from the model’s generative distribution under target property conditions.
- Conclusion: Drug-like molecules sampled from ZINC demonstrated effectiveness in both property prediction and conditional molecular design.The experiments evaluated the approach using drug-like molecules from the ZINC database.
- Conclusion: The model generates novel molecules satisfying target conditions without an extra optimization procedure.Conditional generation is performed by sampling from the learned generative distribution given desired property targets.
- Conclusion: The procedure learns implicit knowledge from data without requiring fully labeled molecular training data.This supports the use of partially annotated datasets for conditional molecular design.
- Conclusion: Generated SMILES structures require further examination to obtain realistic molecules with desired properties.The model’s outputs are candidate structures rather than validated realistic molecules.
- Conclusion: The study covers only the portion of chemical space representable by SMILES, motivating alternatives with broader and more accurate molecular coverage.The authors identify broader chemical-space representation as a direction for improving applicability.