Back

AlphaFold2 and AlphaFold3 leads to significantly different results in human-parasite interaction prediction

Ozden, B.; Cuesta Astroz, Y.; Karaca, E.

2024-09-21 bioinformatics
10.1101/2024.09.19.613643 bioRxiv
Show abstract

MotivationParasitic diseases pose a significant global health challenge with substantial socioeconomic impact, especially in developing countries (Dampier et al., 2009; Mehmood et al., 2017; Merrifield et al., 2016). Combatting these diseases is difficult due to two factors: the resistance developed by parasites to existing drugs and the underfunded research efforts to resolve host-parasite interactions (Cuesta-Astroz & Oliveira, 2018). To address the latter, we focused on 276 human-parasite domain-domain interaction predictions, involving 15 parasitic species known to cause infections in humans (Table S1, Cuesta-Astroz et al., 2019). We aimed to validate these interactions by using all AlphaFold2-Multimer (AF2) and AlphaFold3 (AF3) (Jumper et al., 2021; Abramson et al., 2024). Our initial hypothesis was that the interactions with an AF confidence score [≥] 0.8 could serve as high confident templates for further research. However, through systematic modeling of these interactions, we observed a striking discrepancy: AF3 confidence assignment behaves significantly differently from that of AF2. In this study, we investigate the underlying causes of this difference.

Matching journals

The top 5 journals account for 50% of the predicted probability mass.

1
Bioinformatics
◐ 1204 papers in training set
Top 1%
19.0%
Similar papers in this journal
2
Computational and Structural Biotechnology Journal
● 242 papers in training set
Top 0.1%
9.9%
Similar papers in this journal
3
PLOS Computational Biology
● 1863 papers in training set
Top 4%
8.1%
Similar papers in this journal
4
Bioinformatics Advances
◐ 203 papers in training set
Top 0.5%
7.5%
Similar papers in this journal
5
Protein Science
○ 246 papers in training set
Top 0.5%
6.9%
Similar papers in this journal
50% of probability mass above
6
Scientific Reports
○ 3612 papers in training set
Top 21%
4.4%
7
PeerJ
◐ 308 papers in training set
Top 1%
4.2%
8
Communications Biology
○ 993 papers in training set
Top 4%
3.5%
9
Briefings in Bioinformatics
◐ 354 papers in training set
Top 3%
3.3%
10
Frontiers in Immunology
○ 638 papers in training set
Top 4%
2.5%
11
Proteins: Structure, Function, and Bioinformatics
○ 88 papers in training set
Top 0.5%
2.5%
12
BMC Bioinformatics
○ 457 papers in training set
Top 3%
2.5%
13
Frontiers in Bioinformatics
○ 49 papers in training set
Top 0.4%
1.8%
14
Frontiers in Cellular and Infection Microbiology
○ 109 papers in training set
Top 1%
1.8%
15
Journal of Molecular Biology
○ 232 papers in training set
Top 2%
1.5%
16
Computers in Biology and Medicine
○ 128 papers in training set
Top 3%
1.2%
17
iScience
○ 1154 papers in training set
Top 29%
1.1%
18
PLOS ONE
● 5266 papers in training set
Top 57%
0.9%
19
International Journal of Molecular Sciences
○ 494 papers in training set
Top 14%
0.9%
20
Journal of Chemical Information and Modeling
● 238 papers in training set
Top 2%
0.9%
21
eLife
● 5828 papers in training set
Top 65%
0.9%
22
mSphere
● 302 papers in training set
Top 7%
0.6%
23
Biomolecules
○ 100 papers in training set
Top 3%
0.6%
24
Structure
○ 193 papers in training set
Top 3%
0.6%

"Similar papers" are the closest papers from that journal in the model's embedding space. They show what the match is built on, but the ranking comes mostly from a classifier over the whole training set, not from these examples alone.