PaperlayerView exact revision on arXiv

Source-linked AI summary

Multi-Objective De Novo Drug Design with Conditional Graph Generative Model

Yibo Li, Liangren Zhang, Zhenming Liu

arXiv:1801.07299v3q-bio.QMcs.LG

TL;DR

De novo drug design must search an enormous, discontinuous molecular space, while existing graph generators are often general and computationally expensive. This paper develops a molecule-focused graph generator and a conditional version for multi-objective drug-design tasks, reporting stronger validity and high enrichment rates for compounds meeting specified requirements.

  • Problem

    De novo drug design searches an enormous, discontinuous space of synthetically available molecules, while previous graph generators are often insufficiently optimized and computationally expensive.

  • Method

    The paper proposes a molecule-focused sequential graph generator without atom-level recurrent units and a flexible conditional model for multi-objective molecular generation.

  • Results

    The graph-based model outperforms SMILES-based methods across several metrics, especially valid-output rate, while the conditional model achieves high enrichment rates for compounds satisfying drug-design requirements.

  • Takeaways & Limitations

    Conditional graph generation provides a promising solution for scaffold, drug-likeness, synthetic-accessibility, and multi-target molecular design tasks.

  • Takeaways & Limitations

    The models completely ignore molecular stereochemistry, which the paper identifies as important for drug development and applicability.

Abstract

from arXiv · show

Recently, deep generative models have revealed itself as a promising way of performing de novo molecule design. However, previous research has focused mainly on generating SMILES strings instead of molecular graphs. Although current graph generative models are available, they are often too general and computationally expensive, which restricts their application to molecules with small sizes. In this work, a new de novo molecular design framework is proposed based on a type sequential graph generators that do not use atom level recurrent units. Compared with previous graph generative models, the proposed method is much more tuned for molecule generation and have been scaled up to cover significantly larger molecules in the ChEMBL database. It is shown that the graph-based model outperforms SMILES based models in a variety of metrics, especially in the rate of valid outputs. For the application of drug design tasks, conditional graph generative model is employed. This method offers higher flexibility compared to previous fine-tuning based approach and is suitable for generation based on multiple objectives. This approach is applied to solve several drug design problems, including the generation of compounds containing a given scaffold, generation of compounds with specific drug-likeness and synthetic accessibility requirements, as well as generating dual inhibitors against JNK3 and GSK3$β$. Results show high enrichment rates for outputs satisfying the given requirements.

Introduction

De novo drug design searches an enormous, discontinuous molecular space, motivating learned generative approaches. This work proposes a scalable graph-based generator and applies conditional generation to drug-design objectives.

  • Motivation: Drug design seeks new chemical entities with desirable pharmacological properties, but synthetically available molecules number around 10^60∼10^100 in a discontinuous chemical space.This scale and structure make molecular searching difficult.
  • Prior approaches: Earlier computational approaches generated molecules through atom-based elongation or fragment-based combination, often coupled with global optimization.Examples include ant colony, genetic, and particle swarm optimization.
  • Prior approaches: Deep generative models have effectively modeled SMILES with recurrent networks, language models, and variational autoencoders for molecule generation and focused design.These approaches were applied to tasks including drug-like compound generation and DRD2 activity design.
  • Representation gap: SMILES generation requires learning grammar and atom-ordering rules unrelated to molecular structure, whereas graph generation represents molecules more directly.The paper therefore focuses on direct molecular-graph generation.
  • Contribution: The proposed graph generator covers ChEMBL compounds with up to 50 heavy atoms and outperforms SMILES-based methods across validity, KL and JS divergence, and NLL metrics.A conditional version addresses scaffold, property, and multi-objective drug-design tasks.
  • Graph generative model: The framework sequentially builds molecular graphs from an empty graph by sampling neural-network-parameterized transitions along a decoding route.The decoding policy selects transitions using the generation history until termination.
  • Graph generative model: The model reduces generation cost with a simpler decoding scheme, no atom-level recurrent unit, and randomized route sampling controlled by α.These design choices target improved scalability for molecular graphs.

Decoding Scheme

The decoding scheme generates a molecular graph through four transition types, while the proposed architectures avoid atom-level recurrence. MolMP uses only the current graph state, and MolRNN adds a molecule-level recurrent unit.

  • Decoding Scheme: The transition set contains initialization, append, connect, and termination actions for constructing molecular graphs.Initialization adds the first atom; append adds and bonds a new atom; connect bonds the latest atom to an existing atom.
  • Decoding Scheme: Merging atom addition with its bond formation makes generation require exactly |E| + 2 steps, generally fewer than the corresponding SMILES length.The merged operation is the paper’s main decoding-scheme simplification.
  • Decoding Policy: At each step, the decoding policy outputs append probabilities for atom–bond combinations, connect probabilities for bond types, and a termination probability.These probabilities are computed from the decoding history by a neural network.
  • MolRNN: MolRNN extends MolMP with a single molecule-level recurrent unit while retaining the same overall architecture.Neither proposed architecture uses atom-level recurrence.
  • MolMP: MolMP models each transition as depending only on the current graph, pθ(t|Gi, ..., G0) = pθ(t|Gi), rather than the full generation history.This removes recurrent units and allows computation at different training steps to be parallelized.
  • Network Architecture: Atom representations are initialized from atom type and latest-atom status, then passed through stacked graph-convolutional layers.The architecture concatenates outputs from all layers before subsequent processing.
  • Graph Convolution: Each graph-convolutional layer combines an atom’s previous representation with information from local and distance-conditioned remote neighbors.Remote-neighbor information expands the receptive field with fewer layers.

MolRNN

MolRNN extends MolMP with a single molecule-level recurrent unit that updates the graph representation during sequential generation. This recurrent component improves performance with little additional computational cost.

  • Architecture: MolRNN adds a single molecule-level recurrent unit to the MolMP architecture.The unit updates its hidden state using the latest appended atom and the molecule representation.
  • Architecture: The recurrent network uses three GRU layers with a hidden size of 512.
  • Architecture: MolRNN calculates transition probabilities by replacing the molecule representation hGi with the recurrent hidden state hRNN_i+1.
  • Performance: The molecule-level recurrent unit provides significant performance improvements while inducing little extra computational cost compared with MolMP.

Graph Convolution

The graph convolutional architecture represents each atom using information from both directly connected neighbors and more distant atoms. Bond types condition local-neighbor information, while distances condition remote-neighbor information.

  • Architecture: Each graph convolutional layer adopts a BN-ReLU-Conv structure.
  • Architecture: The receptive field size D is set to 3, incorporating remote neighbors to reach a larger receptive field with fewer layers.
  • Information sources: The output representation of atom v is computed from local and remote neighbor information.
  • Information sources: Local-neighbor information is conditioned on the bond type b between atom v and neighborhood atom u.
  • Information sources: Remote-neighbor information is conditioned on the distance d between atom v and remote neighbor u.

Likelihood Function

Training maximizes molecular-graph log-likelihood, but step-wise generation makes likelihood tractable only for specific decoding routes. The model therefore uses randomized route sampling controlled by α to approximate the marginal likelihood.

  • Likelihood objective: Training maximizes the log-likelihood pθ(G) of the training samples.
  • Likelihood objective: For a given decoding route, the joint log-likelihood is the sum of the log transition probabilities across generation steps.
  • Likelihood estimation: The marginal likelihood sums over all possible decoding routes, but this calculation is intractable for most drug-design molecules.
  • Likelihood estimation: The sampled route distribution qα(r|G) follows depth-first canonical decoding while allowing random mistakes with probability 1−α.
  • Likelihood estimation: For α = 1, route sampling reduces to deterministic decoding, and α is optimized as a model hyperparameter.

Conditional Generative Model

The conditional generative model converts requirements into a code that guides graph generation, supporting scaffold, property, and dual-inhibitor design tasks. Its scaffold workflow uses fingerprints derived from structures extracted from DrugBank.

  • Conditional model: Requirements are converted into a conditional code c, and the decoding policy is conditioned on c.
  • Conditional model: Conditional models can be applied to multi-objective and multi-task settings more easily than fine-tuning-based methods.
  • Conditional implementation: The graph convolution incorporates c as an additional input term, while the SMILES model concatenates c with the input at every step.
  • Applications: The model is applied to scaffold-based generation, property-based generation, and designing dual inhibitors of JNK3 and GSK3β.
  • Scaffold-Based Generation: The scaffold set S is extracted from approved DrugBank drugs using Bemis-Murcko scaffolds and ring assemblies.
  • Scaffold-Based Generation: A binary scaffold fingerprint marks whether each molecule contains each scaffold in S, and the query scaffold fingerprint conditions generation.
  • Scaffold-Based Generation: The scaffold-conditioning procedure provides detailed control over the scaffold of generated structures.
  • Property-Based Generation: Drug-likeness and synthetic accessibility are treated as important requirements in de novo drug design.

Designing Dual Inhibitor Against JNK3 and GSK-3β

The conditional graph generator is applied to designing compounds active against both JNK3 and GSK-3β, using activity predictors and staged training to address sparse positives.

  • Task and conditional representation: Dual-inhibitor generation targets JNK3 and GSK-3β, two serine/threonine kinases linked to disease pathogenesis.The conditional code indicates activity against each target with binary values.
  • Activity labeling: Random forest classifiers trained on ExCAPE-DB activity data assign JNK3 and GSK-3β activity labels to ChEMBL compounds.The classifiers use ECFP6 molecular representations.
  • Data imbalance and training: Only 1.2% of ChEMBL molecules are predicted active against JNK3 or GSK-3β, producing low enrichment during conditioned generation.This imbalance motivates first training the model without conditioning, followed by fine-tuning on the 1.2% subset.
  • Training setup: The graph models are trained on ChEMBL with MXNet and Adam, using an initial learning rate of 0.001 and decay rate of 0.001 every 100 iterations.No model-architecture optimization is performed beyond fitting the model into memory.
  • Comparison models: SMILES baselines include a variational autoencoder, two GRU language models, and an LSTM language model.The GRU baselines use 512 or 1024 hidden units, while the LSTM uses three layers of 1024 units.

Evaluation Metrics

The study evaluates generative models with several metrics to assess their performance and sample quality.

  • Several metrics are employed to evaluate the performance of generative models.
  • The evaluation framework treats generative-model performance as a multi-metric assessment rather than a single-score comparison.
  • Metric-based evaluation is used to assess generative models in the study.

Sample Validity

Sample validity and distributional quality are assessed through validity, novelty, and molecular-property comparisons using divergence measures.

  • Validity and novelty: 300,000 structures are generated per model and evaluated with RDKit for the rate of valid outputs.This tests whether generated structures are chemically correct.
  • Validity and novelty: Novelty is measured as the rate of generated compounds absent from the training set.
  • Property distributions: Generated and test-set distributions of molecular weight, LogP, and QED are compared to assess modeling of molecular properties.
  • Distributional metrics: Kullback–Leibler divergence compares generated and test-set molecular-property distributions.
  • Distributional metrics: Jensen–Shannon divergence is also used to compare generated and test-set molecular-property distributions.
  • Distributional metrics: KL and JS divergence values are computed with SciPy kernel density estimation using a Gaussian kernel and Scott’s Rule bandwidth.

Negative Log-Likelihood

Negative log-likelihood is evaluated on canonically ordered test graphs for graph–SMILES comparison, while conditional models are assessed with control, enrichment, reproduction, and diversity measures.

  • Negative log-likelihood: Negative log-likelihood is evaluated on the test set to measure model performance.
  • Negative log-likelihood: For graph–SMILES comparison, test-set NLL is evaluated using canonical ordering.
  • Negative log-likelihood: Graph-model NLL is reported only for models trained with α = 1, because values for α < 1 cannot be directly compared across models.The study therefore relies more on DKL and DJS for α < 1 models.
  • Conditional-generation metrics: The conditional-control matrix Kcc′ measures how outputs sampled under code c satisfy condition c′.When c = c′, Kcc′ gives the correctly generated-output rate Rc; high Rc and low off-diagonal values indicate control.
  • Conditional-generation metrics: Enrichment over random EORc compares the conditional-generation rate with the rate of molecules satisfying condition c in training data.
  • Target and diversity metrics: Target-based generation additionally reports reproduced-molecule rates, while conditional-output diversity is measured with internal diversity based on Tanimoto distance.The target evaluation uses separate JNK3/GSK-3β condition sets and test-set active molecules.

Model Performance and Sample Quality

The graph-based generators generally produce more valid molecules and show strong modeling performance, while conditional versions achieve enriched outputs across scaffold and property-based tasks. Graph models also expose interpretable structural errors and learned scaffold-specific patterns, though diversity is sometimes lower.

  • Unconditional model performance: MolRNN achieves the best reported NLL at 24.08, while MolMP uses fewer parameters than SMILES GRU1 despite weaker NLL performance.MolMP uses 1 × 10^6 parameters versus 4 × 10^6 for SMILES GRU1.
  • Output validity: Both MolRNN and MolMP outperform all SMILES-based methods in valid-output and valid-and-novel-output rates.Changing α from 1.0 to 0.8 significantly increases validity for both graph-based models.
  • Error patterns: SMILES-based invalid outputs mainly arise from grammar mistakes, whereas graph-based failures mainly involve broken aromaticity during stepwise generation.Atom-valence mistakes are relatively minor for graph-based models.
  • Error patterns: Graph-based outputs are comparatively interpretable because broken aromaticity and unclosed aromatic rings can be corrected through explicit-hydrogen refinement or added aromatic bonds.These corrections may introduce additional bias to the outputs.
  • Property distribution modeling: MolRNN gives the best DKL and DJS performance for molecular weight and QED, while SMILES GRU2 performs best for LogP.Changing α from 1.0 to 0.8 significantly improves MolMP and MolRNN across molecular-property distributions.
  • Conditional generation: Conditional graph models achieve higher validity across scaffold conditions and strong scaffold enrichment, with EORc > 1,000 for scaffolds 1–3 and EORc > 100 for scaffold 4.Graph models perform better on scaffold 3 for Rc, while SMILES models perform better on another reported comparison.
  • Conditional generation: SMILES-based conditional outputs are generally more diverse than graph-based outputs, although the differences are minor relative to standard deviations.The authors suggest this may indicate slight overtraining of the graph-based model.
  • Conditional generation: Scaffold-conditioned generation reproduces scaffold-specific side-chain characteristics, including restricted substitution positions and long aliphatic chains for scaffolds 1–3.The observed patterns reflect reported SAR for Ang II receptor antagonists and may arise from learning dataset bias.

Generating Dual Inhibitors for JNK3 and GSK-3β

The conditional graph generator is applied to generate dual and selective inhibitors for JNK3 and GSK-3β. It achieves strong enrichment and interpretable target-associated structures, while performance varies by selectivity objective and some known actives are reproduced.

  • Task setup: Predictive models label ChEMBL compounds for conditional training with AUC=0.983 for JNK3 and AUC=0.984 for GSK-3β.The labeled dataset is then used to train the conditional graph generator.
  • Comparative performance: Graph models outperform SMILES models in validity for GSK-3β-selective and JNK3-selective compounds, whereas SMILES models perform better for dual inhibitors.The better architecture depends on the selectivity condition.
  • Comparative performance: SMILES models achieve better Rc and EORc for dual inhibitors and GSK-3β-selective inhibitors, while graph models achieve better performance for JNK3-selective inhibitors.Both architectures show high enrichment over random across selectivity combinations.
  • Selectivity outcomes: Dual-inhibitor generation produces many GSK-3β-positive and JNK3-negative outputs, although both models still achieve high enrichment over random for all selectivity combinations.JNK3-selective and dual inhibitors are rarer in ChEMBL than GSK-3β-selective inhibitors.
  • Chemical-space analysis: Generated molecules under different selectivity conditions occupy distinct chemical-space regions and tend to lie near test-set samples in t-SNE projections.The projections use ECFP6 fingerprints and include generated and test-set molecules.
  • Structural interpretation: Generated structures frequently contain established target-associated scaffolds, including diaminopurines and triazolones for JNK3 and 2,3-bis-arylmaleimides for GSK-3β.Aminopyrimidines occur across conditions but are more enriched among dual inhibitors.
  • Recovered actives: The model reproduces 10.3% of JNK3 test-set molecules and 6.0% of GSK-3β test-set molecules despite excluding those test molecules from ChEMBL training.The result indicates generation of confirmed positives not seen during training of the predictive model and conditional generator.
  • Framework scope: The framework uses a less computationally expensive graph convolutional architecture and a flexible approach to decoding invariance for multiple drug-design tasks.The model is trained on ChEMBL compounds and scaled to molecules containing up to 50 heavy atoms.
Loading 1801.07299v3…