How to Use GenMol

GenMol is a generalist molecular design model built around discrete diffusion over SAFE fragment sequences. The paper is notable because one model handles de novo generation, fragment linking, scaffold completion, and optimization by editing fragment-level representations instead of training a separate generator for each medicinal-chemistry task.

On Neurosnap, Mode selects between DeNovo, Onestep, Completion, and Remask. Seeds can be left blank for de novo generation or used to provide starting SMILES for fragment-constrained design and local optimization. Sampling controls such as Temperature, Randomness, and Gamma let researchers decide how conservative or exploratory the run should be.

The workflow fits early hit finding as well as later analog design. Instead of returning only novel structures, GenMol makes it easy to compare proposals by overall optimization score together with medicinal-chemistry indicators such as drug-likeness, synthetic accessibility, and lipophilicity.

The core representation is SAFE, a fragment-based molecular string derived from BRICS decomposition in which molecular identity is not tied to a single left-to-right tokenization. GenMol uses a masked discrete-diffusion process with a BERT-style denoiser, so generation is bidirectional and parallel rather than strictly autoregressive. That is a better fit for fragment rearrangement and substructure-level editing than token-by-token atom strings.

A key innovation in the paper is fragment remasking. Rather than perturbing an entire molecule indiscriminately, the model can mask selected fragments of a seed compound and regenerate only those parts. This is scientifically useful because medicinal chemists often reason in terms of replacing, linking, or decorating fragments rather than rewriting whole molecules from scratch.

On Neurosnap, Number Per Seed, Number of Iterations, Budget, Top K, and the annealing settings provide control over search breadth and optimization pressure. Researchers usually interpret the output as a comparison set: sort candidates by score, inspect how QED, SA, and LogP shift across modes, and decide whether the chemistry looks more suitable for hit expansion, scaffold hopping, or property correction.

• Flexible campaign entry: GenMol supports both seed-free discovery and seed-guided optimization, which covers hit finding and lead refinement in one model family.
• Fragment-level editing: SAFE and fragment remasking make the search easier to interpret than atom-wise mutation strategies.
• Optimization control: Sampling, frontier, budget, and annealing settings let researchers tune novelty, search depth, and exploitation pressure.
• Medicinal-chemistry review: Candidate ranking is naturally paired with QED, synthetic accessibility, and LogP so chemists can triage proposals quickly.