Protein folding by AI: wrinkles

Tech giants Alpha and Meta (Google and Facebook) applied their Artificial Intelligence (AI) to fold proteins computationally, predicting 3-dimensional shapes from the 1-dimensional sequence data. Meta’s paper is still paywalled (preprint) but AlphaFold’s Nature papers from last year are available (Jumper, Tunyasuvunakool). 

The AlphaFold authors noted that the ~100,000 protein structures determined by conventional experimental means are a small portion of the “billions” extant in nature. Previous approaches “focus on either the physical interactions or the evolutionary history”, which they say relies on the availability of close homologues or works for (only) a few, small proteins and is otherwise “computationally intractable” (too hard). They evaluated in the 87 protein domains comprising the 14th Critical Assessment of (protein) Structure Prediction (CASP14) dataset, structures not yet deposited in the public Protein Data Bank (PDB). This permits a ‘blind’ (apriori) comparison of AI methods, by comparing their predictions with the newly-solved structures. 

Fig 1. a. Scores. b. Backbone. c. Side chains

 By this measure, AlphaFold is much better than its competitors (Fig 1a, shown, predicted vs experimental). It gains accuracy on backbone and side chains (1b, c) “by incorporating novel neural network architectures and training procedures based on the evolutionary, physical and geometric constraints of protein structures”. Using “multiple sequence alignments (MSAs) and pairwise” comparisons, it “predicts the 3D coordinates of all heavy atoms for a given protein using the primary amino acid sequence and aligned sequences of homologues as inputs”. So give it a bunch of similar sequences and structures and voila! it gives the ‘new’ one. Thy describe the process and you can download to code to inspect, modify, run yourself (open source).  

While impressive, this is a very constrained set of structures, nothing justifying the claims made in the popular press of solving all proteins. To be comprehensive, it seems that AI will have to consider biology, implement means of including the amino-terminal-first synthesis, nucleation, domain folding, insertion into a membrane, and above all interaction with chaperone proteins. 

Jumper J, Evans R, Pritzel A, Green T, Figurnov M, Ronneberger O, Tunyasuvunakool K, Bates R, Žídek A, Potapenko A, Bridgland A, Meyer C, Kohl SAA, Ballard AJ, Cowie A, Romera-Paredes B, Nikolov S, Jain R, Adler J, Back T, Petersen S, Reiman D, Clancy E, Zielinski M, Steinegger M, Pacholska M, Berghammer T, Bodenstein S, Silver D, Vinyals O, Senior AW, Kavukcuoglu K, Kohli P, Hassabis D. Highly accurate protein structure prediction with AlphaFold. Nature. 2021 Aug;596(7873):583-589. doi: 10.1038/s41586-021-03819-2. Epub 2021 Jul 15. PMID: 34265844; PMCID: PMC8371605.

Leptin boosts protective vaccine responses

Follicular helper T cells (Tfh) support B cell development and production of antibodies, essential for a protective vaccination response. Metabolism has been linked to T cell development and the metabolic hormone leptin varies up to 10-fold among healthy people. Here, the investigators asked whether leptin levels might influence T cell development and contribute to variability in vaccine responses. 

Within a cohort of 76 healthy adults, they found non-responders to influenza vaccination had on average 2.5-fold lower serum leptin levels, with non-responders 10-fold more frequent in the low leptin group (fig 1ab). Tfh counts correlate with leptin levels (fig 2). Similar observations were made among older flu vaccine recipients (age >64 yr) and young Hepatitis B vaccine (HBV) recipients. Adding leptin to T cells cultured in vitro increased Tfh markers and production of IL-21 (fig 2e). 

In mice, they found leptin in areas of B cell development and leptin receptors on Tfh cells. Leptin receptor deficiency reduced antibody responses (fig 3b, c) and (consequently) allowed viral growth (panel a) in mice infected with H1N1 influenza. Tfh in leptin-receptor-deficient mice produced less IL-21 (fig 5b) and supplemental IL-21 restored most antibody production (fig 5a). IL-21 production is abrogated in T cells lacking STAT3 (fig 5g), strongly supporting a mechanism involving STAT3 and IL-21. 

Fig 7. Leptin protects from fasting-induced susceptibility to influenza. 

 

They could transiently reduce serum leptin levels by ‘fasting’ (starving) mice on alternate days 5 to 15 days after infection with influenza (Fig 7a, shown above). This timing chosen to avoid interfering with T cell priming (d 0-5) and focus on peak Tfh development (starting d5). Supplemental leptin protected against influenza (panel b), underscoring the significance of this pathway.

Deng J, Chen Q, Chen Z, Liang K, Gao X, Wang X, Makota FV, Ong HS, Wan Y, Luo K, Gong D, Yu X, Camuglia S, Zeng Q, Zhou T, Xue F, He J, Wei Y, Xiao F, Ma J, Hill DL, Pierson W, Nguyen THO, Zhou H, Wang Y, Shen W, Sun L, Li Z, Xia Q, Qian K, Ye L, Rockman S, Linterman MA, Kedzierska K, Shen N, Lu L, Yu D. The metabolic hormone leptin promotes the function of TFH cells and supports vaccine responses. Nat Commun. 2021 May 24;12(1):3073. doi: 10.1038/s41467-021-23220-x. PMID: 34031386; PMCID: PMC8144586.

Physical activity increases gut bacteria diversity

Previous work established associations in humans between physical activity and reduced obesity, reduced mortality, and improved cardiovascular health. Physical activity has been also associated with the microbiome in animals. Here, the relationship between physical activity and microbiome in humans was investigated.

The authors studied a cohort of 720 adults, citizens of Wisconsin, average age 55 years, 83% White, 10% Black, 42% male. Gut microbial, (bacterial) composition was assessed using sequencing the V3-V4 region of 16S rRNA extracted from stool samples. Note this is only a subset of the ‘microbiome’, not include non-bacterial components such as fungi, viruses, etc. 

They monitored physical activity using accelerometers worn on the hip (activity) or wrist (sleep). Participants also self-reported whether in a typical week they walked or biked at least 10 minutes continuously to get around. Those who responded ‘yes’ were classified as participating in ‘active transportation’. Note this is a threshold of less than half a mile a week, walking only about 100 m per day. 

Table 2. Linear mixed effects models (adjusted for characteristics, Table 1). CI, confidence interval; SD, standard deviation; MVPA, moderate to vigorous physical activity; ** p<0.05. *** p < 0.01.

They identified 865 unique bacterial taxa, largely encompassed by about 20 abundant phyla (Fig 1). They observed no change in bacterial diversity in participants who engaged in moderate-to-vigorous activity (line 2, Table 2, shown) or active transportation (line 3). However, when they analyzed those participants who engaged in higher levels of active transportation, at least 1 standard deviation (SD) above the average, they observed significant increases in bacterial diversity (line 4).

They also found the abundance of an unknown family from order Clostridiales was associated with increased weekly MVPA minutes. They conclude that their results “point to a potential pathway by which the gut micro- biota may be linked to physical activity and other well established health benefits”.

Holzhausen EA, Malecki KC, Sethi AK, Gangnon R, Cadmus-Bertram L, Deblois CL, Suen G, Safdar N, Peppard PE. Assessing the relationship between physical activity and the gut microbiome in a large, population-based sample of Wisconsin adults. PLoS One. 2022 Oct 26;17(10):e0276684. doi: 10.1371/journal.pone.0276684. PMID: 36288361; PMCID: PMC9605031.