MegaMolBART

ANWENDUNG VON KI UND HPCIN DER ARZNEIMITTELFORSCHUNG

UNTERSTÜTZUNG DER DATENGESTEUERTEN FORSCHUNG UND ENTWICKLUNG IN DER PHARMAZEUTISCHEN INDUSTRIE
Supported by deep learning and neural transformer networks.

Transformer-based large language models operate at supercomputing scale with BioNeMo.
Accelerate key drug discovery applications.

Chemoinformatics in detail
Chemoinformatics in detail
Transformer-based large language models create new opportunities for real-time exploration of the chemical universe. BioNeMo is a domain-specific framework based on NeMo Megatron for training and deploying biomolecular LLMs at supercomputing scale. It includes the MegaMolBART, ESM-1b, and ProtT5 transformer models.
MegaMolBART is a generative chemistry model trained with 1.4 billion molecules (SMILES strings) that can be used for a variety of chemoinformatics applications in drug discovery, such as reaction prediction, molecular optimization, and de novo small molecule generation.
For ProtT5 and ESM-1b, it has been shown that unsupervised pre-training can be used to generate learned embeddings that incorporate features for predicting protein structure, function, cellular location, water solubility, membrane binding, conserved and variable regions, and more.
Predict protein structures
Predict protein structures
Deep-learning-based approaches such as RELION enable high-throughput automation of cryoelectron microscopy (cryo-EM) to determine protein structures In RELION, an empirical Bayesian approach to cryo-EM analysis is implemented to make single or multiple 3D reconstructions as well as 2D class average more accurate.
To understand protein structures with atomistic detail, tools such as MELD can be used to infer structures from sparse, ambiguous, or imprecise data. MELD uses data in a physics-based Bayesian framework to improve protein structure determination.


Accelerate virtual screening
Accelerate virtual screening
Support molecular dynamics simulations
Support molecular dynamics simulations
GPU-based molecular dynamics frameworks can simulate the fundamental mechanisms of cells and calculate how strongly a particular drug binds to the desired target protein. Potentials acquired through machine learning that show promise for precision, energies, and forces at the quantum mechanical level fundamentally change molecular simulation.
Clara Discovery incorporates a variety of tools and frameworks for molecular simulation, including GROMACS, NAMD, Tinker-HP, VMD, TorchANI, and DeePMD-Kit.

Discover solutions for accelerated computing
Discover solutions for accelerated computing
Optimized for R&D in the pharmaceutical industry
Optimized for R&D in the pharmaceutical industry
Clara Discovery is designed to run on NVIDIA DGX™ A100, the world's most advanced AI system with five petaFLOPS of performance. DGX A100 is designed specifically for all large-scale accelerated computing workloads, providing researchers with the shortest time to solution and a unified, easy-to-deploy infrastructure to support the next generation of drug discovery.
view HPC Application Performance
view HPC Application Performance
view white paper on pharmaceutical industry PODs
learn more about Schrödinger's advanced computing platform
Learn more about NVIDIA DGX A100
Full stack acceleration

NVIDIA DGX SYSTEMS
Editable editable, click me for edit, editable, click me for edit, editable, click me for edit ...
Editable editable, click me for edit, editable, click me for edit, editable, click me for edit ...
Editable editable, click me for edit, editable, click me for edit, editable, click me for edit ...
UNIVERSAL ARCHITECTURE FOR MEDICAL KI INSTRUMENTS

Editable editable, click me for edit, editable, click me for edit, editable, click me for edit ...
Editable editable, click me for edit, editable, click me for edit, editable, click me for edit ...
Details
Editable editable, click me for edit, editable, click me for edit, editable, click me for edit ...
Editable editable, click me for edit, editable, click me for edit, editable, click me for edit ...
Editable editable, click me for edit, editable, click me for edit, editable, click me for edit ...
Editable editable, click me for edit, editable, click me for edit, editable, click me for edit ...
Editable editable, click me for edit, editable, click me for edit, editable, click me for edit ...
Editable editable, click me for edit, editable, click me for edit, editable, click me for edit ...
Editable editable, click me for edit, editable, click me for edit, editable, click me for edit ...
Editable editable, click me for edit, editable, click me for edit, editable, click me for edit ...
Editable editable, click me for edit, editable, click me for edit, editable, click me for edit ...
Editable editable, click me for edit, editable, click me for edit, editable, click me for edit ...
Komponenten | Funktion |
---|---|
NVIDIA RTX 6000 | Discrete GPU |
Xavier AGX 32 GB DRAM Module | CPU, GPU and I/O Processing |
Mellanox ConnectX-6 Smart Interconnect | Xavier AGX 32 GB DRAM Module |
250GB SSD | Removable Storage |
Komponenten | Funktion | |
---|---|---|
CPU | 8-core Carmel ARM v8.2 64-bit CPU, 8MB L2 + 4MB L3 | |
DRAM | 32GB 256-Bit LPDDR4x | 136.5GB/s | |
GPU | (RTX 6000) | 4608-core Turing GPU with 576 Tensor Cores and 72 RT Cores |
(AGX Xavier) | 512-core Volta GPU with 64 Tensor Cores | |
GPU Memory | (RTX 6000) | 24 GB GDDR6 |
GPU FP32 Performance | (RTX 6000) | 16.3 TFLOPS |
Storage | (AGX Xavier) | 32GB eMMC 5.1 |
(Removable SSD) | 250 GB | |
Power | 350 W | |
Encode | (AGX Xavier) | Up to 2x 1000MP/sec |
(RTX 6000) | Up to 1000MP/sec | |
Decode | (AGX Xavier) | Up to 2x 1500MP/sec |
(RTX 6000) | Up to 1500MP/sec |
Komponenten | Funktion | |
---|---|---|
1 x PCIe Gen4 x8 128 Gbps | (RDMA to GPU) | |
1 x PCIe Gen4 x8 128 Gbps | (to AGX Xavier) | |
2 x USB 3.1 Gen 2 | ||
1 x USB 2.0 | ||
1 x 100 Gbps Ethernet QSFP | (To Mellanox) | |
1 x 10 Gbps Ethernet RJ45 | (To Mellanox) | |
1 x 1 Gbps Ethernet RJ45 | (to AGX Xavier) | |
1 x SD Card | ||
1 x USB-C | Debug Interface | |
Video Input | HDMI In | CSI In |
Video Output | HDMI Out (from AGX Xavier) | DP Out (from RTX 6000) |
The Clara AGX Developer Kit is available exclusively to members of the NVIDIA Clara Developer Partner Program. Register for the Developer Partner Program through the application and an NVIDIA sales representative will contact you to determine next steps. Users can only sign up with a corporate or university email address.
Get the Clara AGX Developer Kit
- NVIDIA RTX support - Docker and CUDA, TensorRT, dGPU activation.
- NVIDIA Rivermax 100GbE Streaming - Rivermax transport protocol streams data over Ethernet directly into GPU GDDR DRAM with GPU Direct technology.
- NVIDIA Clara Guardian Support - Develop AI applications to improve patient care and operational efficiency using everyday sensors such as cameras and microphones.
- Windows device mode support - I/O driver enables Jetson AGX to operate in device mode on Windows 10 machines.
- Endoscopy with NVIDIA DeepStream - Video in and inference for endoscopy and other video-based modalities
- Sensor processing - support for serial camera interfaces
- Reference application for AI endoscopy and ultrasound
ADDITIONAL DEVELOPMENT RESOURCES
Clara Discovery is a growing collection of frameworks, applications and models that enable GPU-accelerated computational drug discovery. In particular, Clara Discovery supports genomics workflows with Clara Parabricks, cryoEM pipelines with Relion, virtual screening with Autodock, protein structure prediction with MELD, various third-party molecular simulation applications, pre-trained models and training frameworks from Clara Imaging, and Clara NLP with BioMegatron, pre-trained models from BioBert, and the NeMo training framework.
Clara Parabricks is a computational framework that supports genomic applications from DNA to RNA. It leverages NVIDIA's CUDA, HPC, AI, and data analytics stacks to build GPU-accelerated libraries, pipelines, and reference application workflows for primary, secondary, and tertiary analyses. Clara Parabricks is a complete portfolio of off-the-shelf solutions coupled with a toolkit to support the development of new applications that meet the needs of genomics labs.
Clara Imaging is a domain-optimized application framework for developers that includes a TensorFlow-based training framework with state-of-the-art pre-trained models to kick-start AI development using techniques such as Transfer Learning, Federated Learning, and AutoML. To enable faster creation of AI-enabled data, Clara Train includes APIs for AI-powered annotations that make any medical viewer AI-enabled.
Clara NLP is a collection of SOTA biomedical pre-trained language models and highly optimized pipelines for training NLP models on biomedical and clinical texts. With NeMo and BioMegatron, researchers and data scientists can build even more powerful NLP models on the large corpus of text data available to them.
Autodock is a growing collection of computational docking and virtual screening methods used in structure-based drug discovery and the study of fundamental mechanisms of biomolecular structure and function.
Clara is based on NVIDIA CUDA 10.1.243, which requires NVIDIA driver version 418.xx. However, if you are working with Tesla (e.g. T4 or another Tesla card), you can use NVIDIA driver version 384.111+ or 410. You can also use driver version 396 for Tesla T4.
The end user license agreement is included with the product. The licenses are also included in the zip file of the model application. By dragging and using the Clara Train SDK container and downloading models, you accept the terms of these licenses.
EINE FLEXIBLE CLOUD-INFRASTRUKTUR FÜR IHR VIRTUELLES RECHENZENTRUM
WARUM sysgen?
we are your partner for complex IT tasks in the software, hardware and medical solutions sector.
In addition to NVIDIA's flagship DGX H100, we naturally also offer other alternatives from the manufacturers. Whether NVIDIA DGX or Supermicro HGX systems in NVIDIA certified servers, single GPUs or industry specific server configurations-
we have the right solution for every budget.