Torso

We inference the five trained models and use the predicted confidence score to select the best model per target.
Using our CASP14 configuration for AlphaFold, the trunk of the network is run multiple times with different random choices for the MSA cluster centres (see Supplementary Methods 1.11.2 for details of the ensembling procedure). The full time to make a structure prediction varies considerably depending on the length of the protein. Representative timings for the neural network using a single model on V100 GPU are 4.8 min with 256 residues, 9.2 min with 384 residues and 18 h at 2,500 residues. These timings are measured using our open-source code, and the open-source code is notably faster than the version we ran in CASP14 as we now use the XLA compiler⁷⁵ .
Since CASP14, we have found that the accuracy of the network without ensembling is very close or equal to the accuracy with ensembling and we turn off ensembling for most inference. Without ensembling, the network is 8× faster and the representative timings for a single model are 0.6 min with 256 residues, 1.1 min with 384 residues and 2.1 h with 2,500 residues.
Inferencing large proteins can easily exceed the memory of a single GPU. For a V100 with 16 GB of memory, we can predict the structure of proteins up to around 1,300 residues without ensembling and the 256- and 384-residue inference times are using the memory of a single GPU. The memory usage is approximately quadratic in the number of residues, so a 2,500-residue protein involves using unified memory so that we can greatly exceed the memory of a single V100. In our cloud setup, a single V100 is used for computation on a 2,500-residue protein but we requested four GPUs to have sufficient memory.
Searching genetic sequence databases to prepare inputs and final relaxation of the structures take additional central processing unit (CPU) time but do not require a GPU or TPU.

The following summarizes the steps required for a developer to turn an existing Rosetta application into a ROSIE server. It is a snapshot of the protocol at the time of writing. A continuously updated version of this protocol is being made available at http://goo.gl/Sh7oB. Importantly, the protocol has been written by new ROSIE developers and so captures the perspective required to promote faster first development cycles for other new engineers. ROSIE development tools and source code are available to registered developers through RosettaCommons.
Download the VM (http://graylab.jhu.edu/ROSIE) and open it with VirtualBox (http://www.virtualbox.org). Before you start, you may want to do a ‘svn update’ in ‘∼/rosie’ and ‘∼/R/trunk/rosetta’, and rebuild the Rosetta trunk, since they may be out of date.
1 Modify the file ‘rosie/rosie.front/development.ini’. Find the line ‘host = 192.168.0.64’ and comment it out. Enable the line ‘host = 127.0.0.1’.
2 To run the server: Open two terminals. In one of them, cd into ‘rosie/rosie.back’ and execute ‘./run_rosie-daemon.sh’. In the other terminal, cd into ‘rosie/rosie.front’ and execute ‘./run-rosie-server.sh’.
3 Open ‘localhost:8080’ in your browser. Login as admin (password: managepass).
1 Create your application in rosie.back/protocols/XXX. You need at least two files: submit.py and analyze.py. See “rna_denovo” for example files.
2 For machine-dependent files, edit rosie.back/data.template/XXX. Edit rosie.back/rosie-daemon.ini.template, add useful shorthands and add the app into the protocol line. Copy the corresponding files to rosie.back/data/XXX and rosie.back/rosie-daemon.ini so the VM server can read the files.
3 Add the corresponding controller in rosie.front/rosie/controllers/XXX.py. See rna_denovo.py as an example.
4 Add your controller into controllers/root.py. In root.py, search for ‘rna_denovo’. Add the two corresponding lines for your application.
5 During the creation of the controller files, you may want to make some validation checks for the input format. They are in rosie.front/rosie/lib/validators. You might need to create your own validation tests.
6 Create your page in rosie.front/rosie/templates/XXX/. You need at least 3 pages: index.html, submit.html, and viewjob.html. See rna_denovo for example.
7 Link your application to the main page in template/index.html.
8 You may want an icon. Put a png file of ∼ 1024*1024 into rosie/public/image/XXX_icon.png, and link it to the pages.
9 For documentation, create pages in template/documentations. Also you need to edit controllers/documentation.py to let the server know where it is. Then link your documentation to documentation/index.html and in the other pages of your application.
10 Edit rosie.front/rosie/websetup/bootstrap.py and add the name of the new app.
11 Go to rosie.front/. Run ‘source ∼/prefix/TurboGears-2.2/bin/activate’ then ‘python update_protocol_schema.py’ to update the database.
12 Test the new application in the browser of the VM to make sure it runs fine.
13 Create a new file rosie/doc/XXX.txt, put a short description of protocol input, output, and command line flags. Also add an example job, with input files and a simple readme, into rosie/examples/validation_tests.
14 Commit the changes (use ‘svn commit –username XXXX’ to specify the user name of the commit). Inform the ROSIE administrators for integration into the central server.

Several sensitivity analyses were used to check and correct for the presence of pleiotropy in the causal estimates. Cochran’s Q was computed to quantify heterogeneity across the individual causal effects, with a P-value ≤ 0.05 indicating the presence of pleiotropy, and that consequently, a random effects IVW MR analysis should be used^{43 (link),48 (link)}. We also assessed the potential presence of horizontal pleiotropy using MR-Egger regression based on its intercept term, where deviation from zero denotes the presence of directional pleiotropy. Additionally, the slope of the MR-Egger regression provides valid MR estimates in the presence of horizontal pleiotropy when the pleiotropic effects of the genetic variants are independent from the genetic associations with the exposure^{49 (link),50 (link)}. We also computed OR estimates using the complementary weighted-median method that can give valid MR estimates under the presence of horizontal pleiotropy when up to 50% of the included instruments are invalid^{44 (link)}. The presence of pleiotropy was also assessed using the MR-PRESSO. In this, outlying SNPs are excluded from the accelerometer-measured physical activity instrument and the effect estimates are reassessed^{51 (link)}. For all of the aforementioned sensitivity analyses to identify possible pleiotropy, we considered the estimates from the extended 10 SNP instrument as the primary results due to unstable estimates from the 5 SNP instrument. A leave-one-SNP out analysis was also conducted to assess the influence of individual variants on the observed associations. We also examined the selected genetic instruments and their proxies (r² > 0.8) and their associations with secondary phenotypes (P-value < 5 × 10⁻⁸) in Phenoscanner (http://www.phenoscanner.medschl.cam.ac.uk/) and GWAS catalog (date checked April 2019).
For the extended 10 SNP instrument, we also conducted multivariable MR analyses to adjust for potential pleiotropy due to BMI because the initial GWAS on physical activity reported several strong associations (P-value < 10⁻⁵) between the identified SNPs and BMI^{52 (link)}. The new estimates correspond to the direct causal effect of physical activity with the BMI being fixed. The genetic data on BMI were obtained from a GWAS study published by The Genetic Investigation of ANthropometric Traits (GIANT) consortium^{53 (link)} (Supplementary Table 9). Additionally, for the extended 10 SNP instrument, we also conducted analyses with adiposity-related SNPs (i.e. those previously associated with BMI, waist circumference, weight, or body/trunk fat percentage in GWAS studies at P-value < 10⁻⁸) excluded (n = 5; rs34517439, rs6775319, rs11012732, rs1550435, rs59499656). Finally, we conducted two-sample MR analyses using BMI adjusted GWAS estimates for the 5 SNP accelerometer-measured physical activity instrument^{11 (link)}. However, the MR results using the BMI adjusted GWAS estimates should be interpreted cautiously due to the potential for collider bias^{11 (link)}.
All the analyses were conducted using the MendelianRandomisation^{54 (link)} and TwoSampleMR^{55 (link)} packages, and the R programming language.

Deconvolution was performed using the CibersortX algorithm at cibersortx.stanford.edu [74 (link)]. Single-cell transcriptomic profiling dataset of cells in the embryonic pancreas [39 (link)] was used as a reference, including count matrix and metadata labels. Particularly, only cells with pancreatic epithelial or endocrine cell fate were used, corresponding to the annotation of five broader cell types—α cells, β cells, endocrine progenitors, trunk epithelium and tip epithelium [39 (link)]. The reference matrix was built out of the 2589 cells and gene list of 18,565 gene features, as deposited by [39 (link)]. Each cell population counted > 250 cells. The units of the reference matrix were UMI counts. Calculation of the scRNA-seq signature matrix was done in default mode (quantile normalization disabled, minimal expression of 0.75, replicates of 5, sampling of 0.5). Imputation of cell fractions and group-mode expression were used in default settings, with S-mode batch correction enabled, quantile normalization disabled and n = 100 permutations for significance analysis. Sample mixture file was submitted with unfiltered gene list 27,124 features for Isl1CKO and in UMI counts.