Table 1 The benchmark results on PCQM4MV2

From: Data-driven quantum chemical property prediction leveraging 3D conformations with Uni-Mol+

Model

# param.

# layers

Valid MAE (↓)

Leaderboard MAE1 (↓)

MLP-Fingerprint15

16.1M

-

0.1735

0.1760

GCN33

2.0M

-

0.1379

0.1398

GIN34

3.8M

-

0.1195

0.1218

GINE-V N3,9,35

13.2M

-

0.1167

-

GCN-V N9,33

4.9M

-

0.1153

0.1152

GIN-V N9,34

6.7M

-

0.1083

0.1084

DeeperGCN-V N3,36

25.5M

12

0.1021

-

GraphGPSSMALL37

6.2M

5

0.0938

-

TokenGT11

48.5M

12

0.0910

0.0919

GRPEBASE12

46.2M

12

0.0890

-

EGT13

89.3M

24

0.0869

0.0872

GRPELARGE13

46.2M

18

0.0867

0.0876

Graphormer4,5

47.1M

12

0.0864

-

GraphGPSBASE37

19.4M

10

0.0858

-

GraphGPSDEEP37

13.8M

16

0.0852

0.0862

GEM-238

32.1M

12

0.0793

0.0806

GPS++39

44.3M

16

0.0778

0.07202

Transformer-M3

47.1M

12

0.0787

-

 

69M

18

0.0772

0.0782

Uni-Mol+

27.7M

6

0.0714   ± 6e − 5

-

 

52.4M

12

0.0696   ±  5e − 5

0.0708

 

77M

18

0.0693  ±  3e − 5

0.0705

  1. 1 The leaderboard was accessed on October 15, 2023, the date of this paper’s submission.
  2. 2 GPS++’s leaderboard submission consists of a 112-model ensemble and utilizes the validation data for training.
  3. We highlight the best results in bold. Source data are provided as a Source Data file.