CLaSMO

A generative model that draws a molecule from scratch can hand a medicinal chemist a structure nobody knows how to make: high-scoring on a property oracle, and unsynthesizable. Real medicinal chemistry works the other way round, starting from a scaffold that already exists and making small local edits, because the starting point is known to be makeable and the change is small enough to reason about.

CLaSMO (Conditional Latent Space Molecular Scaffold Optimization) builds that editing loop into a generative framework. A molecule is decomposed with BRICS, which cuts only at bonds retrosynthetic rules allow to be broken and reformed, and the resulting fragments train a conditional variational autoencoder. The condition is the atomic environment of the attachment point — atom type, hybridization, valence, formal charge, degree and ring membership of the atom that will carry the new bond — so the decoder proposes fragments that can actually bond where you put them. Latent Space Bayesian Optimization then searches jointly over the latent vector and the choice of bonding atom, and every candidate must clear a Dice similarity floor before the expensive oracle is called.

Onur Boyar, Hiroyuki Hanada and Ichiro Takeuchi developed it at Nagoya University, Takeuchi also leading a team at the RIKEN Center for Advanced Intelligence Project. It was posted as a preprint in November 2024 and published in Transactions on Machine Learning Research in September 2025.

#Key Features

• Attachment-site conditioning: A small autoencoder compresses the six atomic features at each BRICS division point into a condition vector that enters both the encoder input and the latent space, steering the decoder toward chemically compatible fragments.
• Joint search over fragment and position: A product kernel — an RBF over the continuous latent vector times a Kronecker-delta kernel over the discrete bonding atom — lets one Upper Confidence Bound acquisition pick what to add and where.
• Similarity as a hard constraint: Candidates below the Dice similarity threshold on Morgan fingerprints are penalized rather than scored, keeping the output recognizable as a derivative of the starting compound.
• Human-in-the-loop web application: A public Streamlit app runs the frozen checkpoint on user SMILES and lets a chemist pick the region to modify rather than leaving it to the search.

#Technical Details

The CVAE is trained once on QM9's roughly 130,000 small organic molecules, which BRICS turns into 18,706 unique substructure-plus-environment pairs. Fragments are one-hot SELFIES matrices over an 18-symbol alphabet, encoded by three fully connected layers and decoded by three GRU layers. The latent space is deliberately tiny: two dimensions, still reaching over 90% reconstruction accuracy on held-out fragments, and a size Bayesian optimization handles well. Where a chemical language model such as MoLFormer-XL is pretrained on 1.1 billion SMILES strings, CLaSMO's generator is compact and the search does the heavy lifting.

Evaluation uses the 20 tasks of the Gao et al. (2022) sample-efficiency benchmark — rediscovery, multi-property optimization, structural hopping and activity oracles such as DRD2 and GSK3β — with 100 oracle evaluations per molecule rather than the benchmark's usual 10,000, over 10 seeds and 100 starting scaffolds from QM9 and ZINC250K. Against Graph-GA, Smiles-GA, STONED, MolDQN, MARS and SynFlowNet, CLaSMO takes the best Top-10 average score in 13 of 20 tasks at a similarity threshold of 0.50, 13 of 20 at 0.25, and 11 of 20 unconstrained. Edited molecules average 3.73 on synthetic accessibility against 3.00 for the input scaffolds, and a retrosynthesis classifier judges 88.0% of them synthesizable. Removing the conditioning drops the Top-10 QED score from 0.930 to 0.897, below random sampling from the conditional latent space at 0.922.

#Applications

CLaSMO is built for lead optimization where each measurement is expensive — a docking campaign, an enzymatic assay, a wet-lab synthesis — and a project can afford tens rather than thousands of evaluations. A team with a hit compound can ask for the substituent additions most likely to raise potency or drug-likeness while keeping the molecule close enough to the original that the existing synthetic route still applies. The web application extends this to chemists with no machine learning workflow, who point at the position they want decorated and review each proposed modification.

#Impact

CLaSMO argues a specific case: under a tight oracle budget, a small conditional generator with a well-designed search outperforms unconditioned models that need far more evaluations. Putting the similarity constraint inside the optimization objective rather than filtering afterwards is what makes it usable for scaffold work, and reporting synthetic alongside retrosynthetic accessibility sets a higher bar than property scores alone. Distribution is modest: the encoder, decoder and condition-embedding checkpoints sit in the GitHub repository rather than a model hub, the repository carries no license file, and the hosted application caps each run at 10 iterations, making it a demonstration rather than the full method.