GenMol: A Drug Discovery Generalist with Discrete Diffusion
Authors: Seul Lee, Karsten Kreis, Srimukh Prasad Veccham, Meng Liu, Danny Reidenbach, Yuxing Peng, Saee Gopal Paliwal, Weili Nie, Arash Vahdat
ICML 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We experimentally validate Gen Mol on a wide range of molecule generation tasks that simulate real-world drug discovery problems, including de novo generation, fragment-constrained generation, goal-directed hit generation, and goal-directed lead optimization. Across extensive experiments, Gen Mol outperforms existing methods by a large margin (Figure 1). |
| Researcher Affiliation | Collaboration | 1KAIST. Work during an internship at NVIDIA. 2NVIDIA. Equal advising. Correspondence to: Seul Lee <EMAIL>. |
| Pseudocode | Yes | Algorithm 1 Goal-directed Molecular Optimization of Gen Mol Input: A set of molecules D, vocabulary size V , decomposition rule for fragment vocabulary Rvocab, decomposition rule for fragment remasking Rremask, number of generations G Set F fragments obtained by decomposing D with Rvocab Set V top-V fragments of F (Eq. 5) Set plen fragment length distribution of D based on Rremask Set M while |M| < G do Select and attach two fragments from V to get xinit Sample the fragment length m plen Select one of the fragments of xinit based on Rremask and replace it with m mask tokens to get xmask Generate xnew by iteratively unmasking xmask with Gen Mol Update M M {xnew} Decompose xnew into fragments {f 1,f 2, . . . } with Rvocab Update V top-V fragments from V {f 1,f 2, . . . } end while Output: Generated molecules M |
| Open Source Code | No | The paper does not contain an explicit statement about releasing the source code for Gen Mol, nor does it provide a link to a code repository for their methodology. It mentions using code for baselines (JT-VAE and Di Gress) but not for their own proposed model. |
| Open Datasets | Yes | Gen Mol is trained on the SAFE dataset (Noutahi et al., 2024), which combines molecules from ZINC (Irwin et al., 2012) and Uni Chem (Chambers et al., 2013). ... To construct an initial fragment vocabulary, we adopted a simple decomposition rule Rvocab that randomly cut one of the non-ring single bonds of a given molecule three times and apply it to the ZINC250k dataset. |
| Dataset Splits | Yes | Table 6: Statistics of the SAFE dataset. Train Test Validation Number of examples 945,455,307 118,890,444 118,451,032 |
| Hardware Specification | Yes | Gen Mol was trained using 8 NVIDIA A100 GPUs. The training took approximately 5 hours. All the molecular generation experiments were conducted using a single NVIDIA A100 GPU and 32 CPU cores. |
| Software Dependencies | No | The paper mentions using several software libraries such as Hugging Face Transformers, RDKit, and Therapeutics Data Commons (TDC) but does not provide specific version numbers for any of them. |
| Experiment Setup | Yes | We used the BERT (Devlin et al., 2019) architecture... we set max position embeddings to 256. ... We set the batch size to 2048, the learning rate to 3e 4, and the number of training steps to 50k. We used the log-linear noise schedule of Sahoo et al. (2024) and the Adam W optimizer (Loshchilov & Hutter, 2019) with β1 = 0.9 and β2 = 0.999. ... Gen Mol can balance between quality and diversity by adjusting the values of the softmax temperature τ and the randomness r. ... We used N = 1. We performed the grid search with the search space τ {0.5, 0.8, 1, 1.2, 1.5} and r {1, 1.2, 2, 3}... |