Datasets
Models
34 lines (27 with data), 2.5 kB
Pretrained Molecular Representations
This page contains benchmarks of property prediction models with pre-training.
We have two main methods for the pre-training.
- Self-supervised pre-training is learning the graph structural information in aself-supervised manner. Here we are pre-training on a subset of 2m molecules from`ZINC15`_.
- Supervised pre-training is doing pre-training on a large supervised dataset.Here we are using 456k molecuels and 1,310 tasks from `ChEMBL`_.
For the downstream tasks, we consider the scaffold splitting for molecule data. The split for train/validation/test sets is 80%:10%:10%. For each pre-training method and downstream dataset, we evaluate with 10 random splits and report the mean and the derivation of AUROC metric.
| `BBBP`_ | `Tox21`_ | `ToxCast`_ | `SIDER`_ | `ClinTox`_ | `MUV`_ | `HIV`_ | `BACE`_ | Avg. | |
|---|---|---|---|---|---|---|---|---|---|
| No Pretrain | 67.1(2.9) | 75.0(0.2) | 60.6(0.7) | 58.9(0.8) | 60.8(3.9) | 64.3(3.4) | 76.4(1.6) | 66.5(9.0) | 66.2 |
| `InfoGraph`_ | 68.9(0.6) | 76.4(0.4) | 71.2(0.6) | 59.8(0.7) | 70.3(4.2) | 69.4(0.8) | 75.5(0.7) | 73.7(2.6) | 70.7 |
| `EdgePred`_ | 67.1(2.6) | 74.6(0.7) | 69.8(0.5) | 59.4(1.5) | 59.0(2.6) | 66.8(1.0) | 76.3(2.0) | 68.4(3.9) | 67.7 |
| `AttrMasking`_ | 65.2(0.9) | 75.8(0.5) | 70.6(0.6) | 58.9(0.9) | 79.0(2.3) | 68.3(2.1) | 76.9(0.9) | 78.1(0.8) | 71.6 |
| `ContextPred`_ | 71.1(1.8) | 75.6(0.3) | 71.1(0.3) | 61.7(0.5) | 65.9(1.9) | 68.5(0.6) | 77.1(0.3) | 78.6(0.5) | 71.2 |