Geforce gtx 1080

The method was implemented using the popular Python library PyTorch on NVIDIA’s GeForce GTX 1080 with 11 GB of available memory. For evaluating CapsNet-MHC, we applied five-fold cross-validation at the training step. To this end, the training set is split into five nearly equal parts, where four parts are used for the training, and the remained part is used for the model evaluation. To ensure that all parts of the datasets are included in evaluating the proposed model, the training and evaluation phases are repeated in five iterations. Finally, the model is evaluated using an independent test set to demonstrate its prediction capability for any unseen data. In this manner, the average prediction accuracy of various iterations is considered as the final result. Supplementary Note 7 and Supplementary Table S9 illustrate various parameter settings for CapsNet-MHC.

Molecular Dynamics simulations were performed on an AMD 3970 × CPU@ 3.7 GHz with the support of a NVIDIA GeForce GTX 1080 graphics chip using GROMACS v2018 and the CHARMM36 force field built in the Linux Ubuntu 18 environment^{40 (link),41 (link),48 (link)}. Calculations are based on the GRLBD crystal structure with PDB ID 5NFP, solved by Hemmerling et al.^{39 (link)} The L773P mutation was generated with FoldX while dexamethasone topology was generated with the CGenFF server^{38 (link),49 (link),50 (link)}. The Avogadro program was used to assign the dexamethasone hydrogen atom coordinates^{51 (link)}. The Cgenff_charmm2gmx.py script from the Mac Kerell lab was used to format the ligand topology for GROMACS. The unit cell was defined as a dodecahedron and was solvated in TIP3P water. The protein net charge was neutralized by adding the appropriate Na⁺ and Cl⁻ ions. The energy minimization step and NVT, NPT equilibration steps were performed based on the tutorials and mdp. files provided by Dr. Justin Lemkul (http://www.mdtutorials.com/) with minor modifications⁵² . Production MD for data collection was performed for 100 ns (n = 2) and trajectories were analyzed with the GROMACS toolset. Histograms were visualized with Origin 8.6, hydrophobic interactions with BIOVIA discovery studio and Kyte-Doolitle hydrophobicity was colored using UCSF Chimera^{37 (link),53 (link)}.

Implementation of the BME criterion and the optimization framework was written in Python using Jax (Bradbury et al. 2018 ) and Optax (Babuschkin et al. 2020 ). Optimization was performed on a Xeon 2.30 GHz (CPU; Intel Corporation) or on a single GeForce GTX 1080 (GPU; Nvidia Corporation). Evaluation of the BioNJ (Gascuel 1997 (link)) and FastME (Lefort et al. 2015 (link)) methods was performed via the R package ape (Paradis et al. 2004 (link)) using rpy2 (Gautier and Krassowski 2021 ). Tree manipulation and visualization scripts were written using ete3 (Huerta-Cepas et al. 2016 (link)) and NetworkX (Hagberg et al. 2008 ). An implementation is available at: https://github.com/Neclow/GradME.

The HTC Vive virtual reality headset and a TP Cast wireless adapter (latency < 2 ms) are used to display the virtual environment with the following specs: a resolution of 2,160 × 1,200 pixels (1,080 × 1,200 pixels per eye), a field of view of 110° and a refresh rate of 90 Hz. We also use two Vive controllers and three Vive Trackers to track the body movements using inverse kinematics algorithms to animate the avatars in real-time (Figure 4). The computer running the application is composed of an Intel Xeon E5-1607 @ 3,10 GHz processor and a Nvidia GeForce GTX 1080 graphics card.
We measured the overall latency of the system using the manual frame counting method described in He et al. (2000 ). Thanks to a high-speed camera (240 Hz), we recorded simultaneously five hand-clapping motions by removing the lenses of the HTC Vive to be able to compare the delay between the real movement and its virtual counterpart displayed on the headset's screen. Analyzing the motions frame by frame, we measured an average shift of 3 frames (12.5 ms). Thus, our system benefits from these low latency devices providing an overall delay under the threshold of 20 ms recommended for virtual reality (Raaen and Kjellmo, 2015 ).

The VR system used was the HTC VIVE, with the HMD and hand trackers. The HMD has a resolution of 2,160 × 1200 (1,080 × 1,200 per eye), a refresh rate of 90 Hz, a field of view of about 110°and weight 0.47 Kg. The headphones used were the Sennheiser HD 206 headphones. The computer used had an Intel i7 4790k processor, 16 Gb of RAM and a NVidia GeForce GTX 1080 graphics card.

For computational processing, Python 3.6.4 programming language was used in Microsoft Windows operating system. The deep learning model was constructed by using Keras 2.2.2 which utilizes Tensorflow-GPU 1.6.0 as a backend. An NVIDIA GeForce GTX1080 (8 GB RAM) with 16 GB system RAM workstation was used.