Table 3 AbBFN baseline performance can be further improved with rapid fine-tuning

From: Protein sequence modelling with Bayesian flow networks

Method

Data

AAR (%)

 

Train/Test Overlap

Fine-tuned

CDR-H1

CDR-H2

CDR-H3

LSTM†86

× 

✓

41.0 ± 5.2

28.5 ± 1.6

15.7 ± 0.9

C-LSTM†44

× 

✓

40.9 ± 5.4

29.2 ± 1.1

15.5 ± 1.2

RefineGNN†54

× 

✓

39.4 ± 5.6

37.1 ± 3.1

21.1 ± 1.6

C-RefineGNN†44

× 

✓

33.2 ± 3.0

33.5 ± 3.2

18.9 ± 1.4

MEAN†44

× 

✓

58.3 ± 7.3

47.2 ± 3.1

36.4 ± 3.1

AntiBERTy52

✓

× 

76.7 ± 5.3

71.1 ± 5.9

42.7 ± 2.6

AbLang253

✓

× 

76.3 ± 5.7

70.6 ± 4.5

42.7 ± 2.5

AbBFN

× 

× 

70.3 ± 5.4

64.9 ± 4.5

31.5 ± 2.3

AbBFN+

× 

✓

77.1 ± 4.2

72.6 ± 4.1

39.7 ± 2.6

  1. Amino Acid Recovery (AAR) rates for various methods on the 10-fold SAbDab benchmark (presented as mean ± SD). The highest AAR for each task is highlighted in bold. Methods labelled with the symbol † report results from ref. 44. ‘Train/Test Overlap’ report if a model’s pre-training data was cleaned to prevent data leakage into the SAbDab benchmark. AbBFN is found to provide leading performance among methods that avoid data leakage. For these models that have not seen the test distribution, ‘Fine-tuned’ reports if they were subsequently fine-tuned on the nine train folds associated with each test fold. AbBFN+ is obtained using 1000 steps of fine-tuning, and reports, or approaches, leading performance across all CDR regions.
Back to article page