InVirtuoGen

Fragment-based drug discovery works by keeping a known active substructure fixed and rebuilding the chemistry around it. Most sequence-based generative models are a poor match for that workflow: SMILES strings linearize a molecular graph by depth-first traversal, scattering chemically meaningful fragments across the string and giving the model little direct control over what is preserved. InVirtuoGen, from InVirtuoLabs and the Max Planck Institute of Biochemistry, attacks this by generating fragmented SMILES — sequences of BRICS-derived fragment blocks with explicit attachment points — under a discrete flow formulation.

The core methodological claim is a shift from completion to refinement. Autoregressive chemical language models emit tokens left to right, and masked diffusion models progressively unmask positions; in both cases a token, once written, is final. InVirtuoGen instead starts from a uniform source distribution over the entire 204-token vocabulary and lets every position change at every denoising step. Because the training loss scores predictions at all positions rather than only masked ones, the number of sampling steps is decoupled from sequence length — a shorter molecule can be refined more, not less, within the same compute budget.

Released as an arXiv preprint on 30 September 2025 by Benno Kaech, Luis Wyss, Karsten Borgwardt, and Gianvito Grasso, the model covers de novo generation, fragment-constrained design, target-property optimization, and structure-based lead optimization. It sits alongside SMILES-based chemical language models such as MoLFormer-XL on the sequence side, and 3D pocket-conditioned generators such as MODA on the structure side.

#Key Features

• Uniform-source discrete flow: Generation begins from random tokens rather than mask tokens, so all sequence positions are jointly revised at each step and sampling cost is independent of molecule length.
• Fragment-level control: Molecules are encoded as rBRICS fragment blocks with numbered attachment points, extending the SAFE notation, which makes scaffold retention and fragment assembly explicit rather than emergent.
• Bidirectional backbone: A diffusion transformer with full self-attention models long-range dependencies between fragments, suiting the inherently unordered fragment representation.
• Hybrid optimization scheme: A genetic algorithm supplies high-scoring parent molecules as flow starting states, while Proximal Property Optimization — PPO adapted to discrete flows via Monte Carlo log-probability estimation — provides gradient-guided refinement under oracle rewards.
• Single hyperparameter configuration: Unlike baselines tuned per task, one setting is used across all 23 Practical Molecular Optimization tasks and all fragment-constrained subtasks.

#Technical Details

The backbone is a diffusion transformer with 36 layers, 12 attention heads, hidden dimension 768, and rotary positional embeddings; a 12-layer variant is used for the optimization experiments to reduce memory and runtime. Pretraining runs for a single epoch on ZINC and UniChem — roughly one billion molecules, matching the corpora used by SAFE-GPT and GenMol — with AdamW at learning rate 1e-4, batch size 300, and length bucketing capped at 25,000 tokens per batch. Sequence length is sampled separately from a length prior taken from ZINC250k.

On the PMO benchmark, InVirtuoGen reaches a summed top-10 AUC of 18.99 ± 0.22 across 23 tasks with ZINC250k prescreening, ahead of GenMol (18.36) and f-RAG (16.93); without prescreening it scores 16.68 ± 0.26 versus 16.21 for Genetic GFN, 14.71 for Mol GA, and 14.18 for REINVENT. On five fragment-constrained tasks it averages 29.5 quality and 90.4% uniqueness against GenMol's 28.1 and 82.2%, though at lower validity (71.0% vs 95.4%). In lead optimization across parp1, fa7, 5ht1b, braf, and jak2, summed docking scores reach -152.4 versus -148.7 for GenMol at a 0.4 Tanimoto similarity constraint, with the margin widening at the stricter 0.6 threshold (-145.7 vs -117.7).

#Applications

The model targets the practical stages of a small-molecule campaign: enumerating diverse, synthesizable, drug-like starting points; growing or linking fragments around a fixed motif for linker design, scaffold decoration, and superstructure assembly; and driving multi-objective property and docking-score optimization under drug-likeness and synthetic-accessibility constraints. Because the representation is fragment-native, medicinal chemists can specify what to keep in the same terms they already use, and the sample-efficiency focus of the PMO evaluation — a 10,000-oracle-call budget — reflects settings where scoring is expensive.

#Impact

InVirtuoGen demonstrates that uniform-source discrete flows are competitive with, and on oracle-budgeted optimization better than, the masked-diffusion and autoregressive formulations that dominate fragment-based molecular generation. Pretrained checkpoints are distributed through Hugging Face and Zenodo and the reproduction code is public, though both code and weights carry non-commercial terms that exclude use by or on behalf of commercial organizations. The authors are candid about scope: the fragmented SMILES representation discards stereochemistry, rBRICS decomposition can miss chemically relevant cut points, QED and synthetic-accessibility scores are coarse proxies, no ADMET assessment is included, and every result is computational and awaits experimental validation. The work is a preprint and has not yet been peer reviewed.