Table 3 Ablation study for model backbone and for sampling strategies for q, on PCQM4MV2

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

No.

TriangularUpdate

OuterProduct

Pair Repr.

R

w1.0

w0.0

wu

Valid MAE (↓)

1

✓

✓

✓

1

0.1

0.8

0.1

0.0696

Ablation study on model backbone

2

✗

✓

✓

1

0.1

0.8

0.1

0.0704

3

✗

✗

✓

1

0.1

0.8

0.1

0.0710

4

✗

✗

✗

1

0.1

0.8

0.1

0.0709

5

✓

✓

✓

0

0.1

0.8

0.1

0.0715

6

✓

✓

✓

2

0.1

0.8

0.1

0.0695

7

✗

✗

✗

0

0.1

0.8

0.1

0.0738

Ablation study on training strategy

8

✓

✓

✓

1

1.0

-

-

0.0771

9

✓

✓

✓

1

-

1.0

-

0.1122

10

✓

✓

✓

1

-

-

1.0

0.0724

11

✓

✓

✓

1

0.1

0.9

-

0.0697

12

✓

✓

✓

1

-

0.9

0.1

0.0753

13

✓

✓

✓

1

0.1

0.7

0.2

0.0698

14

✓

✓

✓

1

0.2

0.7

0.1

0.0703

15

✓

✓

✓

1

0.1

0.6

0.3

0.0702

16

✓

✓

✓

1

0.2

0.6

0.2

0.0706

17

✓

✓

✓

1

0.3

0.6

0.1

0.0714

18

✓

✓

✓

1

Noisy Nodes

0.0760

19

✗

✗

✗

1

Noisy Nodes

0.0798

  1. R refers to the number of rounds of conformation updates. w1.0 refers to the sample probability of RDKit conformation, w0.0 refers to the sample probability of target conformation with noise and wu refers to the sa probability of intermediate conformation. We highlight the best results in bold. We use underlines to indicate the results under the standard settings. Source data are provided as a Source Data file.