Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Scalable and Cost-Efficient de Novo Template-Based Molecular Generation
Authors: Piotr Gaiński, Oussama Boussif, Andrei Rekesh, Dmytro Shevchuk, Ali Parviz, Mike Tyers, Robert A. Batey, Michał Koziarski
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We benchmark recent reaction-based GFlow Nets across three library sizes and three design tasks, demonstrating that SCENT significantly outperforms prior methods in terms of synthesis cost, molecular diversity, and the number of high-reward compounds discovered. |
| Researcher Affiliation | Academia | 1 Jagiellonian University, Faculty of Mathematics and Computer Science, 2 Mila Québec AI Institute, 3 Université de Montréal, 4 University of Toronto, 5 The Hospital for Sick Children Research Institute, 6 Acceleration Consortium, 7 Vector Institute, * Equal contribution EMAIL |
| Pseudocode | Yes | The complete pseudocode of SCENT s training procedure is provided in Appendix A.1. |
| Open Source Code | Yes | Our code is publicly available at https://github.com/koziarskilab/SCENT. |
| Open Datasets | Yes | MEDIUM: Syn Flow Net s templates with 64k Enamine building blocks, and (3) LARGE: like MEDIUM, but with 128k building blocks. These settings span practical use cases: SMALL focused on rapid synthesis and MEDIUM/LARGE supporting broader exploration requiring external procurement. The models are tested in the s EH proxy task [7], along with the GSK3β and JNK3 tasks [44 47]. |
| Dataset Splits | No | We evaluated models in three settings: (1) SMALL: curated fragments (see Appendix E) for rapid synthesis or high-throughput screening, (2) MEDIUM: Syn Flow Net s templates with 64k Enamine building blocks, and (3) LARGE: like MEDIUM, but with 128k building blocks. These settings span practical use cases: SMALL focused on rapid synthesis and MEDIUM/LARGE supporting broader exploration requiring external procurement. |
| Hardware Specification | Yes | In Table 6 report GPU runtimes (in minutes) for all template-based baselines on the s EH proxy, averaged over 3 random seeds using a V100 32GB GPU. |
| Software Dependencies | No | All forward policies were trained using the Trajectory Balance objective [31] and their parameters were optimized using Adam [52]. All methods were trained using their built-in action embedding mechanisms and on a maximum number of reactions equal to 4. |
| Experiment Setup | Yes | All the models sampled 320,000 forward trajectories during the training in total in SMALL and MEDIUM settings and 256,000 in LARGE. All forward policies were trained using the Trajectory Balance objective [31] and their parameters were optimized using Adam [52]. All methods were trained using their built-in action embedding mechanisms and on a maximum number of reactions equal to 4. Hyperparameters were chosen semi-manually or using small grid searches to maximize the number of high-reward modes in the SMALL setting in the s EH proxy. We set the number of sampled forward trajectories to 64, and the number of trajectories sampled from the prioritized replay buffer to 32. We set the β to 8 for s EH, and 48 for GSK3β and JNK3. We used uniform ϵ-greedy exploration with ϵ = 0.05 for all our experiments, except for those with other exploration techniques (e.g. Exploitation Penalty). Decomposability Guidance ... αD = 5. Synthesis Cost Guidance ... α = 5. Dynamic Building Block Library We update the dynamic building block library every T = 1000 iterations for a maximum of Nadd = 10 additions (that is, we do this until the end of training). The number of molecules added to the library every time it is updated is L = 400. Exploitation Penalty We set ϵ = 3 in Equation (12) and schedule γ in two ways: 1) so that it linearly decays to zero after N iterations, and 2) it grows with the trajectory length (each step increases it by some constant factor γ). We set N = for SCENT that uses Dynamic Library and N = 1000 for the rest of the runs. The growing delta γ = 0.2, while the initial s0 temperature γ0 = 1.0. |