A Model to Search for Synthesizable Molecules

1 Introduction

• Existing machine-learning approaches generate molecules with desirable properties, but do not describe how to synthesize them for experimental testing.Prior work generated drug-like, biologically active, and soluble molecules.Source: §1 · p. 1
• MOLECULE CHEF selects commonly available reactants and predicts their reaction products, providing a route for making generated molecules.The model retains a continuous latent space for optimization while mirroring a multistage molecule-generation process.Source: §1 · p. 1, Figure 1 · p. 2
• The model addresses practical synthesis concerns, including reaction stability and toxicity, and proposes reactants for target molecules that may be more practical to manage.Source: §1 · p. 2

2 Background

• The background section frames traditional computational discovery and machine-learning approaches as context for identifying issues the paper addresses.Source: §2 · p. 2

2.1 Virtual Screening

• Virtual screening enumerates combinations of building blocks, but its untargeted search covers only a tiny portion of estimated drug-like chemical space.The estimated drug-like space is ∈[10^23, 10^100], while chemical space usually screened is ∈[10^7, 10^10].Source: §2.1 · p. 2

2.2 The Molecular Search Problem

• De novo design searches iteratively for molecules with desired properties, while early discrete methods remained local and later SMILES, tree, graph, and reinforcement-learning approaches addressed search or validity.The cited SMILES CVAE experiment produced valid molecules only 0.7% of the time.Source: §2.2 · pp. 2–3

2.3 The Molecular Recipe Problem

• Property-directed molecular search lacks synthesis pathways, leaving the practical makeability of generated molecules unknown; synthesis planning instead recursively deconstructs molecules using reversed reaction predictors.Source: §2.3 · p. 3

2.4 This Work

• MOLECULE CHEF jointly addresses molecular search and synthesis planning by mapping continuous-space representations to available reactants and then predicting products from reactions.The authors identify reaction-based generation as a source of inductive bias for extrapolation, semantically valid molecules, and synthesis routes.Source: §2.4 · p. 3

3 Model

• MOLECULE CHEF maps latent vectors to bags of easily procurable reactants, which a reaction predictor transforms into product molecules.The reactant pool R is a subset of valid molecular graphs.Source: §3 · pp. 3–4
• The model uses a WAE objective that matches aggregate encodings to a standard normal prior while retaining a VAE-style reconstruction cost.Unlike matching each encoding to the prior, the regularizer matches the marginal distribution over datapoints; the authors report avoiding the VAE’s local-optimum issues.Source: §3 · p. 4

3.1 Encoder and Decoder

• The encoder embeds each reactant with a GGNN, sums embeddings for an order-invariant bag representation, and parameterizes a Gaussian distribution over z.The encoder uses four-layer GGNNs to form graph representations.Source: §3.1 · p. 4, Figure 2 · p. 5
• The RNN decoder sequentially selects reactants from the fixed pool or halts, feeding each selected reactant into the next step.Reactants may occur multiple times in a bag, although many practical reactions use only one instance of a particular reactant.Source: §3.1 · pp. 4–5, Figure 3 · p. 5

3.2 Adding a predictive penalty loss to the latent space

• A two-hidden-layer predictor estimates final-product QED from the reactant-bag latent encoding for local optimization.QED is used as a surrogate because it can be deterministically calculated from a molecule with RDKit.Source: §3.2 · p. 5

4 Evaluation

• The training data contains 34,426 unique reactant bags and 4,344 unique reactants, retained when each reactant appeared in at least 15 training reactions.The dataset is based on the USPTO data processed by Jin et al., with reagents filtered out.Source: §4 · p. 6

4.1 Generation

• Generation is assessed using validity, uniqueness, novelty, normalized quality, and FCD; most prior metrics are extended to bags of products.Sampling from the prior yielded single-product bags 97% of the time.Source: §4.1 · p. 6
• Generated products are often valid and frequently pass quality filters, although the reaction predictor can make invalid edits.Table 1 reports validity, uniqueness, novelty, normalized quality, and FCD for samples from the model prior.Source: §4.1 · p. 7, Table 1 · p. 7

4.2 Local Optimization

• Purposeful latent-space optimization improves the overall distribution of best-found QED scores compared with random walks.Figure 4 also shows the starting-location distribution sampled from training data.Source: §4.2 · p. 7, Figure 4 · p. 8

4.3 Retrosynthesis

• The retrosynthesis setup regresses products to latent space, then decodes that representation into a bag of reactants.The evaluation reconstructs a product by applying the Molecular Transformer to the predicted reactant bag.Source: §4.3 · pp. 7–8, Figure 5 · p. 8
• Predicted reactants do not always match the complete correct bag, but often produce structures similar to the target product.The examples use USPTO test-set products that can be made in one step from the predefined reactant pool.Source: §4.3 · p. 8, Figure 6 · p. 8
• Reconstructed-product QED scores correlate with original-product scores, more strongly for reachable products than unreachable products.The model is trained on reachable product reactions; some unreachable products may require reactants outside its pool, at least for one-step reactions.Source: Figure 8 · p. 8, §4.3 · p. 9

4.4 Qualitative Quality of Samples

• A domain expert judged many CVAE- and GVAE-generated molecules to have undesirable properties, unlike MOLECULE CHEF samples.The expert assessed stability, toxicity, oxidizing power, and corrosiveness.Source: §4.4 · p. 9, Figure 9 · p. 9

5 Discussion

• MOLECULE CHEF generates molecules from products of one-step reactions using a pool of predefined reactants.The model selects reactants and runs chemical reactions while optimizing in a continuous latent space.Source: §5 · p. 9
• The approach combines strengths of VAE-based models with classical discrete de-novo design based on virtual reactions.Source: §5 · p. 9
• Future work proposes larger reactant vocabularies and multi-step reactions to widen the range of generated molecules while retaining synthesis-route suggestions and often semantically valid outputs.Source: §5 · p. 9

A.1 Generation Benchmarks on ZINC

• Table 2 reports validity, uniqueness, novelty, and normalized quality for baseline models trained on ZINC.The four metrics are percentages, with higher values better.Source: Table 2 · p. 14, §A.1 · p. 14
• The table's generation metrics are calculated from 20k molecules decoded from random samples of the prior p(z).Source: Table 2 · p. 14
• Normalized quality measures the proportion passing Brown et al.'s quality filters, scaled so the USPTO-derived training dataset scores 100.Source: Table 2 · p. 14
• FCD measures distance between generated molecules and the USPTO-derived training dataset used in the main paper.The caption identifies FCD as Fréchet ChemNet Distance.Source: Table 2 · p. 14

A.2 Further Random Walk Examples and Rationales for Expert Annotation

• Figure 9 labels flag unstable enols and aminals, reactive radicals, unstable ring systems, and toxic thioketones or sulfur-chloride bonds.Source: §A.2 · p. 14
• Figure 10 rationales identify unstable gemthiolol and gem-aminohydroxyl structures and corrosive acyl fluoride, chloride, and bromide groups.Source: Figure 10 · p. 14
• Figure 11 rationales flag toxic or explosive hydrazines, unstable hemithioacetals and hemiacetals, and unfavorable or antiaromatic ring systems.The annotations also note a potentially oxidizable structure and possible toxicity associated with an N-N bond.Source: Figure 11 · p. 15

A.3 Further Retrosynthesis Results

• The section provides additional retrosynthesis examples.Source: §A.3 · p. 15
• An extra experiment assesses whether the pipeline finds molecules with similar properties without reconstructing the correct reactants.Source: §A.3 · p. 15

• Figure 12 shows predicted reactants for products outside the training data when the reactants belong to MOLECULE CHEF’s vocabulary.Source: Figure 12 · p. 16
• Figure 13 shows predicted reactants for products outside the training data when at least one reactant is outside MOLECULE CHEF’s vocabulary.Source: Figure 13 · p. 16

• The reconstruction comparison measures Euclidean distance between ChemNet embeddings after products pass through the retrosynthesis pipeline and Molecular Transformer.Source: §A.3.2 · p. 17
• The random-pair baseline matches each dataset molecule, including reactants and products, with another randomly chosen molecule before measuring embedding distance.Source: §A.3.2 · p. 17
• The product-to-reconstruction distance distribution has greater mass at smaller Euclidean distances than the random-pair baseline.Figure 14 shows the distribution of distances between ChemNet embeddings of the original and reconstructed products.Source: §A.3.2 · p. 17, Figure 14 · p. 17

A.4 Details about our Dataset

• The vocabulary contains 4,344 reactants occurring at least 15 times in USPTO training data, with 34,426 unique co-occurring reactant bags, each associated with a product.Source: §A.4 · p. 17
• Retrosynthesis tests use filtered USPTO test reactions split into reachable cases with all reactants in-vocabulary and unreachable cases with at least one out-of-vocabulary reactant.Reactions with the exact same reactant/product multisets tuple in MOLECULE CHEF training data are filtered out first.Source: §A.4 · pp. 17–18

Details of MOLECULE CHEF’s architecture and parameters

• The encoder sums graph-neural-network molecule vectors to parameterize a 25-dimensional Gaussian latent distribution, while the decoder sequentially selects reactants with a GRU until HALT.The shared gated graph neural network uses four propagation steps and 101-dimensional node representations projected to 50 dimensions.Source: Details of MOLECULE CHEF’s… · pp. 18–19
• A latent-space property predictor estimates QED for local optimization and retrosynthesis, with its loss added to the WAE loss.Source: Details of MOLECULE CHEF’s… · p. 20

Implementation Details for the Baselines in Section 4.1 of Main Paper

• The LSTM baseline uses a 203-item alphabet of bracketed atoms and SMILES symbols; the authors report little qualitative or quantitative performance difference from the 47-item alphabet.The CGVAE includes element-charge-valence triplets occurring at least 10 times and starts generation from a randomly selected node.Source: Implementation Details for the… · p. 20

Other Details

• The authors had not tried a wide range of MOLECULE CHEF hyperparameters and selected a 25-dimensional latent space and learning-rate decay based on validation reconstruction error.They initially tried latent dimensionality 100 and found slightly improved performance after decaying the learning rate by a factor of 10 after 40 epochs.Source: Other Details · p. 20