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.

When viruses cohabit: Flu + RSV = Hybrid Frankenvirus

Experts worry that this winter might be made miserable by unwelcome visitors: something new, coronavirus variants, something flu (influenza A virus, IAV) and something blew, respiratory syncytial virus (RSV). What happens when somebody hosts IAV and RSV at the same time?

These investigators infected cultured human lung cells, A549 cells, and confirmed previous reports that coinfection reduces RSV but not IAV replication (Fig 1). Despite producing lower titer, they observed that infection with IAV appeared to increase the rate of coinfection by RSV.

Fig 3b. Filament with features of IAV and RSV.

IAV and RSV are enveloped viruses that bud from the cell membrane with characteristic glycoproteins hemagglutinin (HA) and fusion (F), respectively. Having detected HA in areas of RSV budding from coinfected cells, the authors hypothesized that some virions would contain components of both viruses. Indeed, they observed many filaments, typical of RSV, with proteins from both viruses, albeit segregated (Fig 2a-e). A remarkable scanning electron micrograph appears to show hybrid viral particles (HVP) budding from the filaments (2f, red arrows). They analyzed the hybrid buds using cryo-ET and were able to ‘segment’ features of both viruses (shown, Fig 3b): mostly IAV virions budding from mostly RSV filaments. 

Amazing biology, but what does it mean clinically? The authors found that the hybrid virions contained IAV capable of infecting cells that had been depleted of their sialic acids, which bind HA, by treatment with neuraminidase (NA), Fig. 4-5). This could be an important mechanism widening the range of infected cells.

Haney J, Vijayakrishnan S, Streetley J, Dee K, Goldfarb DM, Clarke M, Mullin M, Carter SD, Bhella D, Murcia PR. Coinfection by influenza A virus and respiratory syncytial virus produces hybrid virus particles.  Nat Microbiol. 2022 Oct 24. doi: 10.1038/s41564-022-01242-5. Epub ahead of print. PMID: 36280786.

Bad news bears on life choices (vaccine hesitancy)

Vaccine hesitancy, a reluctance or refusal to be vaccinated, probably began when Jenner invented vaccination over two centuries ago. Vaccines have largely eliminated scourges such as smallpox and polio and greatly reduced the rates of other infectious diseases including influenza. Anti-vaccination (anti-vax) stances stem from small, well-established risks of side-effects (managed by a compensation program) and big, vague worries about unrelated, even disproven associations with other maladies. COVID-19 vaccines were developed rapidly and rushed into production, potentially raising valid safely concerns.  However, any valid concerns were allayed when the COVID-19 vaccines were tested and proven safe and effective (Walsh 2020). 

The cable television show Fox News Channel (FNC) amplified concerns about COVID-19 vaccines and downplayed their benefits.  This study of viewership and vaccination covered ~2,750 counties (out of ~3,000 total) in 47 (of 50) US states documents that FNC viewers refused COVID-19 vaccination more often than the viewers of its competitors Microsoft-National Broadcasting Company (MSNBC) or Cable News Network (CNN) (Figure 2, shown).

Figure 2. Effect of network viewership on weekly vaccination rates, 2021

A key question is whether FNC influenced its viewers to refuse vaccination (a cause) or rather were anti-vax viewers attracted to FNC’s messaging, a consequence of playing to its audience.  The investigators used positions in cable channel listing as ‘exogenous shifters’ of viewership (Martin & Yurukoglu 2017).  Viewers are induced into watching more or less of a channel by variation in its position up or down the listing (Fig S3).  They found that “exogenously higher FNC viewership due to channel position causes lower vaccine uptake”.  They show that hesitancy was raised by FNC but not by competitors MSNBC or CNN (Fig 1). Moreover, resistance to vaccination against COVID-19 but not seasonal flu… causal…. Using the channels’ position in the guides.

Their “results imply that watching one additional hour of [FNC] per week for the average household reduces the number of vaccinations by 0.35–0.76 per 100 people”, which would account for a lot of ‘excess deaths’ in many households. Not surprising when “vaccine bad” was said so much more often on FNC than the other channels (Fig S7)! Although they found that FNC’s influence was mostly on those under 65 years old, who are at lower risk severe disease, those younger people are reservoirs of virus for infecting older people. Data-driven lawyers representing survivors of FNC victims could bring class action lawsuits. 

Pinna, M., Picard, L. & Goessmann, C. Cable news andCOVID-19 vaccine uptake. Sci Rep 12, 16804 (2022). 

Lipid metabolism and dementia

About half the human brain mass is lipid. Several brain disorders are known to be caused by abnormal lipid metabolism, disruptions in the processes of making lipids and breaking them down. Second only to Alzheimer’s in prevalence is frontotemporal dementia (FTD), one form of which encompasses a range of social, behavioral, or language disorders (as opposed to memory or motor deficits seen in other disorders).  Several genes have been associated with FTD, foremost among them the conserved genes MAPT (tau), PSEN1 (presenilin), VCP (valosin containing protein) and GRN (granulin).

Granulins are a family of secreted, glycosylated peptides (A, B, C, etc.) cleaved from a single precursor (progranulin, PGRN), that are involved in a wide range of activities probably due to their roles in regulating protein lysosomal protein metabolism.  The authors of this study found that gangliosides (i.e., glycosphingolipids with attached sialic acids) are elevated in brains of granulin mutant mice (GRN R504X), which are analogous to the most prevalent granulin mutation in humans, R493X (substitution of the arginine normally at position 493 with a nonsense codon, resulting in a truncated protein). This mutation causes neuronal ceroid lipofuscinosis, a severe neurodevelopmental disease, in humans and neuroinflammation in mice (Jax).  The metabolic order in the ganglioside degradation pathway (Fig. 1a) is first disialylated GD1 (Fig. 1b, shown, rightmost plot, annotated with red #1) 👉 monosialylated GM1 (#2)  👉 GM2 (#3) 👉 GM3. Also, GD2 👉 GM3 (#3’) via an alternative pathway. Levels of precursor GD1 (#1) are significantly elevated in the brains of mice with heterozygous mutant granulin (Figure 1b: blue fill, Grn +/R493X) compared with normal granulin (Grn+/+, grey) but not in homozygous mutants (purple). This may suggest a feed-back mechanism that limits the accumulation of that metabolite. GM1, #2, is elevated in homozygous mutant brain compared to normal, with the heterozygous mutant intermediate. However, GM2, #3, is not significantly elevated in mutation-bearing mouse brains. The alternate pathway, #3’, shows elevated GD3, the precursor to GM3, in homozygous mutant brains.

They also analyzed the lipids in postmortem human brains of 12 GRN mutation-related FTD cases, 6 sporadic FTD cases, and 3 control normal subjects.  GD3 and GM1 are significantly elevated in GRN-related FTD cases (Fig 1c, blue columns). However, they both seem also elevated in non-GRN (sporadic) FTD cases (green). GD1 is significantly elevated in GRN-related but not -unrelated FTD.  

In a striking simplification, they tested the effects of removing the granulin precursor protein gene, PGRN (same as GRN), in HeLa cells (Fig 2). They found elevated GM2 in the deficient line (GRN-/-) that was reduced to normal levels by restoring granulin (GRN-/- + PGRN-addback).

What causes changes in the levels of gangliosides? Gangliosides are catabolized by lysosomal enzymes. However, those enzymes were not altered by GRN deficiency (Fig 3). An intermediate metabolite, bis(monoacylglycero)phosphate (BMP), which is crucial to ganglioside degradation, was found to be reduced by 50-60% in GRN-deficient HeLa cells and mouse brains, and ‘markedly’ in human brains of FTD cases, although again both GRN-related and sporadic cases (Fig 4d). The authors propose a model wherein “lysosomal granulin peptides maintain lysosomal function and homeostasis, including the levels of BMP, that are crucial for ganglioside catabolism”. Their results await confirmation by others and many details remain to be pursued further. One relatively simple aspect will be clarifying how the model accounts for autosomal dominance of GRN deficiency. Also worth noting is the proximity of GRN and the Alzheimer- and Parkinson-associated MAPT genes, within a million nucleotides in band 17q21.31. 

Boland S, Swarup S, Ambaw YA, Malia PC, Richards RC, Fischer AW, Singh S, Aggarwal G, Spina S, Nana AL, Grinberg LT, Seeley WW, Surma MA, Klose C, Paulo JA, Nguyen AD, Harper JW, Walther TC, Farese RV Jr. Deficiency of the frontotemporal dementia gene GRN results in gangliosidosis. Nat Commun. 2022 Oct 7;13(1):5924. doi: 10.1038/s41467-022-33500-9. PMID: 36207292; PMCID: PMC9546883.

COVID restrictions – Moderation is good

In efforts to reduce the spread of COVID-19, billions of people around the world were subject to rules and laws governing their behavior. Among the various non-pharmaceutical interventions (NPIs), what worked? 

 

An influential early report analyzing data from 11 European countries during the first 4 months of the pandemic found that lockdowns reduced transmission rates (Rt values) significantly, ~80%, whereas other NPIs such as cancelling public events, school closure, encouraging social distancing, and self-isolation, resulted in less significant reductions (0-20%) (Fig 2, Flaxman).  The dramatic drop in infections after lockdown is so obvious that it required no modeling (Fig 1).  A similar study found the most effective NPIs for lowering cases were travel restrictions, school closures, and the partial lockdown (Cortis).  A related study of 19 NPIs during seasonal flu found that banning large gatherings was most effective in limiting transmission (Qiu).  These studies used epidemiological models that directly involve underlying mechanisms.

 

Data on cases, deaths, vaccinations, and tests, were obtained from the COVID-19 Data Hub (Guidotti).  NPI data were obtained from the Oxford Covid-19 Government Response Tracker (OxCGRT).  

 

In this report, the author analyzed data from 132 countries between Feb 2020 to April 2021, capturing 3 waves of infection, beginning March 202, July 2020, and January 2021.  An econometric model with 4 equations: C = cases growth rate, D= deaths growth rate, M = mobility, and p(SI) = probability of the assigned stringency intensity level was employed.  Stringency correlates inversely with nonresidential mobility (Fig 3).  He found that ‘unobserved variables’ influence the growth of cases and deaths (C and D) as well as the stringency (SI) of government policies. Medium-stringency measures greatly reduced case and death growth rates but, surprisingly, yet-more-stringent measures slightly increased them (Fig. 4, shown).   

Fig. 4 Case (panel a) and Death (panel b) growth rates vs NPI Stringency Index. 

Testing helped but contact tracing did little (Fig 5).  The benefits of reduced nonresidential mobility were outweighed by increased within-household transmission.  Various differences in culture, compliance, and enforcement of government imposed NPIs were acknowledged but not clearly managed. Even low levels of vaccination reduced Case and Death growth rates with nonlinear improvement anticipated. 

The findings disagree with previous reports that lockdowns are effective, concluding that "very stringent NPIs provide no further benefits over moderately stringent ones, and that less stringent NPIs function primarily as signals for significant voluntary changes in citizens’ behavior.". Such analyses are crucial for designing effective, results-driven policies and for persuading people to comply.  

 

Spiliopoulos L. On the effectiveness of COVID-19 restrictions and lockdowns: Pan metron ariston. BMC Public Health. 2022 Oct 1;22(1):1842. doi: 10.1186/s12889-022-14177-7. PMID: 36183075; PMCID: PMC9526209.

Antibody specificities are Gut reactions

Germ-free (GF) mice have undeveloped immune systems and practically no antibodies.  How do microbes in the gut (gastro-intestinal tract) stimulate immunity?  This group looked at the specificity of antibodies that develop after GF mice are colonized by individual bacterial species and strains (monocolonization).  They used 8 strains of bacteria to inoculate GF mice and, after 3 weeks, analyzed the specificity of antibodies produced in the gut (recovered from fecal matter) and blood, focusing on the IgA isotype that protects mucosal surfaces.  

Not surprisingly, the antibodies tend to bind specifically to those strains of microbes against which they were stimulated (shown, Figure 1, panel A). They also found that mice monocolonized from birth produce more IgA reactive with that species (termed ‘self’) than newly-introduced species. They make a point about IgA being able to ‘aggregate pathogenic bacteria’ and ‘selectively coat disease-associated bacteria’ but it is unclear how IgA itself could distinguish dangerous from benign and anyway they tested only benign bacteria.  

A small panel (29) of monoclonal IgA antibodies cloned from gut tissues of monocolonized mice also showed species specificity. Finally, they showed that monoclonal IgA antibodies with specific binding activity could be detected in the feces of mice that had been force-fed the IgA 3 hours previously, suggesting a targeted, potential therapy (e.g., against the human pathogen Clostridioides difficile). 

Yang C, Chen-Liaw A, Spindler MP, Tortorella D, Moran TM, Cerutti A, Faith JJ. Immunoglobulin A antibody composition is sculpted to bind the self gut microbiome. ScienceImmunology. 2022 Jul 15;7(73) 

Covid Vax vs Variants

The first RNA vaccines against Covid were based on the Surface, S or “Spike” protein from the reference genome of SARS-CoV-2, Wuhan-Hu-1, published in January 2020.  Since then, more-infectious variants have emerged; one, D614G, dominated but did not escape vaccine protection probably because its characteristic mutation is located outside the Receptor Binding Domain (RBD, 319-541). In the past year, delta variants were overtaken by omicron variants, which have numerous changes in the RBD, raising concern that they escape current RNA vaccines. 

Neutralization assays measure the ability of blood-borne antibodies to block viral infection of cells grown in a culture dish; they are thought to provide valid measurements of protection. These authors previously tested Covid neutralization with blood sera from 15 health care workers, 4 vaccinated with Moderna (mRNA-1273) and 11 with Pfizer-BioNTech (BNT162b2). They confirmed that a third dose (second booster) increased the neutralization activity (titer) against all strains, albeit 3-4 times weaker against subvariants compared with the ancestral version. Moreover, they showed the neutralization titers induced by 3 doses of the vaccines, and therefore presumably levels of protection provided, approximated those found in convalescent (sick) Covid patients (Fig 1 panel B vs C and D). 

In this letter, they studied 46 health care workers, 24 vaccinated and boosted with Moderna and 22 with Pfizer-BioNTech. Fourteen of the cohort were infected during the follow-up year. Figure S3 in the supplementary data confirm the protection provided by a third dose (second booster). They found that neutralization titers against all strains declined over time.  Titers from vaccinated people remained within a substantial fraction of those from infected people (shown, dashed vs solid lines). 

How much protection -- neutralization titer -- is enough? A paper published last year analyzing 7 different vaccines reported that neutralization titers are predictive of protection.  Neutralization titers of 20% of convalescent levels protected on average half of people against detectable infection; levels at 3% convalescent protected against severe Covid. Taken together, these results seem to support the value of the ‘old’ vaccines against the newer variants. 

Qu P, Faraone JN, Evans JP, Zheng YM, Yu L, Ma Q, Carlin C, Lozanski G, Saif LJ, Oltz EM, Gumina RJ, Liu SL. Durability of Booster mRNA Vaccine against SARS-CoV-2 BA.2.12.1, BA.4, and BA.5 Subvariants. N Engl J Med. 2022 Sep 7. 

Obesity and severe COVID

Early in the COVID-19 pandemic, it was recognized that obesity seemed to be related to respiratory failure, or severe acute respiratory syndrome (SARS). These researchers tested whether leptin, a cytokine produced by fat cells in the gut and working on brain cells to influence hunger, might be correlated with risk, perhaps a biomarker of risk. 

They compared 31 obese COVID patients on ventilators with 8 non-infected, non-obese critically ill patients.  They found much higher levels of leptin in the patients with COVID (averages 21 vs 6 ug/l, with very good statistical significance, p = .0007).  The individual measurements overlap (shown), so leptin is not ‘the’ biomarker but clearly related. Whether related as a cause or consequence only a prospective trial could test rigorously. 

Figure 2 detail: BMI of patients panel A and Leptin levels panel B.  In each panel, ‘control’ critically ill patients, left, COVID patients right. 

 

The authors hypothesized that elevated leptin causes a ‘hyper immune’ state, especially stimulating lung epithelial cells.  They noted similar observation previously published for influenza and MERS. Since this paper was published, several groups have reported similar findings 2020 paper that was largely replicated (see review, which cites 3 later papers).  Ironically, shutdowns that have been effective in protecting many people from infection have also increased sedentary lifestyles, BMI, and risk. 

 

van der Voort PHJ, Moser J, Zandstra DF, Muller Kobold AC, Knoester M, Calkhoven CF, Hamming I, van Meurs M. Leptin levels in SARS-CoV-2 infection related respiratory failure: A cross-sectional study and a pathophysiological framework on the role of fat tissue. Heliyon. 2020 Aug;6(8):