UniMatch: Universal Matching from Atom to Task for Few-Shot Drug Discovery
Authors: Ruifeng Li, Mingqian Li, Wei Liu, Yuhua Zhou, Xiangxin Zhou, Yuan Yao, Qiang Zhang, Hongyang Chen
ICLR 2025 | Venue PDF | Archive PDF | Plain Text | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Our experimental results demonstrate that Uni Match outperforms state-of-the-art methods on the Molecule Net and FS-Mol benchmarks, achieving improvements of 2.87% in AUROC and 6.52% in AUPRC. Uni Match also shows excellent generalization ability on the Meta-Mol Net benchmark. The code is available at https://github.com/Lirain21/Uni Match.git |
| Researcher Affiliation | Academia | 1 Zhejiang University 2 Zhejiang Lab 3 Shanghai Jiao Tong University 4 University of Chinese Academy of Sciences 5 ZJU-Hangzhou Global Scientific and Technological Innovation Center |
| Pseudocode | Yes | Algorithm 1 Meta-training procedure for Uni Match. Input: The few-shot training tasks {Tτ}Nt τ=1 of molecular property prediction; Output: Trained model fθ,w; |
| Open Source Code | Yes | The code is available at https://github.com/Lirain21/Uni Match.git |
| Open Datasets | Yes | Our Uni Match outperforms state-of-the-art methods on both the Molecule Net (Section 4.1) and FS-Mol (Section 4.2) benchmarks, achieving improvements of 2.87% in AUROC and 6.52% in AUPRC, respectively. Additionally, we evaluate the generalization ability of Uni Match on the Meta-Mol benchmark, where it demonstrates outstanding performance (Section 4.3). |
| Dataset Splits | Yes | Following the procedural framework of Wang et al. (2021), we adopt AUROC (the area under the receiver operating characteristic curve) as the evaluation metric and set the support set size at 20 (i.e., 2-way 10-shot). The model is trained using the Adam optimizer (Kingma & Ba, 2014). During testing, results are averaged from 10 repeated experiments with different random seeds. For each task, we employ unbalanced sampling to create an uneven distribution of positive and negative samples within the support set. The evaluation metric, AUPRC, is used to effectively assess the model s ability to improve minority class prediction, which is critical in imbalanced datasets (see Appendix C.3). During testing, we set five different sizes for the support set: 16, 32, 64, 128, and 256. For each size, we perform 10 repeated random splits of the support/query sets for the test tasks under these settings and take their averages as the final results. |
| Hardware Specification | Yes | All experiments are run on an NVIDIA RTX A6000 GPU. |
| Software Dependencies | Yes | We implement Uni Match in Py Torch (Paszke et al., 2019) and Pytorch Geometric library (Fey & Lenssen, 2019). |
| Experiment Setup | Yes | We employ the Adam optimizer (Kingma & Ba, 2014) with a learning rate of 0.001 for meta-learning, while using a higher learning rate of 0.05 for fine-tuning the matching module and fusion module within each task. The dropout rate is maintained at 0.1 for all components, except for the graph-based molecular encoder. We summarize the hyperparameters used by Uni Match in Table 3. On FS-Mol benchmark (Stanley et al., 2021), we set the batch task 21 and weight decay 5e-5. And we train the model for 10,000 epoches. On the Meta-Mol Net benchmark, we set the query set size to 8 and the support set size to 2. We employ the Adam W optimizer (Loshchilov & Hutter, 2017) with a learning rate of 0.001 for meta-learning and an inner learning rate of 0.001 for fine-tuning the task-specific modules within each task. A weight decay of 5e-4 is applied. The model is trained for 100 epochs to ensure robust performance. |