Dynotx trains graph neural networks, protein language models, and diffusion-based generative models on GPU clusters to discover novel therapeutics — from in silico screening to preclinical candidates, compressed from years to months.
Our platform integrates graph neural networks for molecular property prediction, transformer-based protein language models, and diffusion models for de novo molecular generation. Each model is trained on large-scale biochemical datasets and deployed through a unified inference stack optimized for NVIDIA GPU architecture.
The pipeline begins with molecular inputs from the ZINC and ChEMBL databases — 2.4 million drug-like compounds represented as atom-bond graphs with 9-dimensional node features (atomic number, charge, hybridization, aromaticity) and 5-dimensional edge features (bond type, ring membership, conjugation).
A graph neural network (Dynotx-MolGNN) performs message passing over the molecular graph to predict properties: binding affinity, toxicity, solubility, and bioavailability. In parallel, a protein language model (Dynotx-ProtLM, based on the ESM-2 architecture) encodes target protein sequences into structural and functional embeddings.
The diffusion model (Dynotx-Diffuse) generates novel 3D molecular structures conditioned on the target protein binding pocket. Generated candidates are docked, scored, and filtered through ADME prediction — producing ranked drug candidates ready for wet-lab validation.
Our end-to-end pipeline processes molecular datasets through GPU-accelerated training, TensorRT-optimized inference, and clinical deployment. Each stage is designed for scale — 2.4M molecules screened, 12,000 GPU-hours per month, 3 drug candidates in preclinical evaluation.
Molecular datasets from ZINC15, ChEMBL, and Protein Data Bank. Standardized SMILES, 3D conformers generated via RDKit + OpenMM. Protein structures prepared with AlphaFold2 predictions and experimental PDB data.
Distributed training on H100 SXM clusters. FSDP sharding across 8 GPUs, bf16 mixed precision, gradient checkpointing. Separate training runs for each model architecture with shared data pipeline.
Models exported to TensorRT and served via Triton. Virtual screening of 2.4M compounds against target proteins. Diffusion model generates novel candidates conditioned on binding pocket geometry.
Top candidates validated through ADME prediction, toxicity filtering, and docking score ranking. Promising hits advance to wet-lab validation. 3 candidates currently in preclinical evaluation across oncology and infectious disease targets.
# Dynotx-ProtLM v2.4 — distributed training configuration # Cluster: 8× NVIDIA H100 80GB SXM · NVLink · InfiniBand model: architecture: esm2_transformer params: 650_000_000 layers: 33 hidden_dim: 1280 attention_heads: 20 max_sequence_len: 1024 positional_encoding: rotary training: gpu_cluster: 8x H100 SXM 80GB strategy: FSDP # Fully Sharded Data Parallel batch_size: 64 # per-GPU, grad accumulation = 4 global_batch: 2048 learning_rate: 4.0e-4 warmup_steps: 8000 total_steps: 250_000 gradient_checkpointing: true mixed_precision: bf16 optimizer: AdamW weight_decay: 0.01 scheduler: cosine_decay data: source: UniRef50 + Pfam + internal sequences total_sequences: 45_000_000 masking: span_masking # 15% masked, avg span 3 augmentation: sequence_crop, random_truncation compute: estimated_gpu_hours: 14_400 peak_memory_per_gpu: 68GB # FSDP sharded interconnect: NVLink 4.0 + InfiniBand HDR framework: PyTorch 2.4 · PyG · CUDA 12.5 inference: runtime: tensorrt server: triton_inference_server quantization: fp16 latency: 47ms # per protein, 512 AA avg throughput: 1200 seq/s
Each model addresses a distinct stage of the drug discovery pipeline — molecular property prediction, protein understanding, and de novo drug design. All models are trained, validated, and deployed independently, then composed into a unified screening pipeline.
| Model | Architecture | Parameters | Training | Performance |
|---|---|---|---|---|
|
Dynotx-MolGNN
dynotx-molgnn-v2.1
Graph neural network for molecular property prediction. Predicts binding affinity, toxicity, solubility, and bioactivity from atom-bond graph representations. 8 message-passing layers with residual connections and edge-conditioned convolutions.
|
GNN (DMPNN)
8 message-passing layers d=512, 9 node feats 5 edge feats edge-conditioned conv |
12.4M
FP16 inference ONNX export PyG backend |
2,800 GPU-hrs
8× H100 SXM batch=512 bf16 mixed precision |
|
|
Dynotx-ProtLM
dynotx-protlm-v2.4
Protein language model based on ESM-2 architecture. Learns structural and functional representations from 45M protein sequences. Used for target identification, binding site prediction, and protein function annotation across the drug discovery pipeline.
|
ESM-2 Transformer
33 layers · 20 heads d=1280 rotary positional encoding span masking (15%) |
650M
FP16 inference Triton serving FSDP training |
14,400 GPU-hrs
8× H100 SXM global batch=2048 250K training steps |
|
|
Dynotx-Diffuse
dynotx-diffuse-v1.3
3D equivariant diffusion model for de novo molecular generation. Generates novel drug-like molecules conditioned on target protein binding pocket geometry. 200 denoising steps with SE(3)-equivariant message passing for 3D coordinate generation.
|
SE(3) Diffusion
200 denoising steps 3D equivariant GNN pocket-conditioned 3D coordinate output |
85M
FP16 inference batch=64 ONNX export |
6,400 GPU-hrs
8× H100 SXM batch=64 classifier-free guidance |
Our models are trained on NVIDIA H100 SXM clusters with FSDP sharding, bf16 mixed precision, and gradient checkpointing. The full stack — from data curation to TensorRT inference — runs on our internal compute infrastructure. We are scaling to 12,000 GPU-hours per month and need additional compute capacity to support expanded molecular screening and protein engineering campaigns.
We evaluate our models on established benchmarks across molecular property prediction (MoleculeNet), protein structure prediction (CAS-14), and molecular generation quality (GuacaMol). All numbers are from single-model evaluation without ensembling.
| Benchmark | Metric | Baseline | Dynotx | Delta |
|---|---|---|---|---|
| MoleculeNet (Tox21) | AUROC | 0.812 | 0.914 | +12.6% |
| MoleculeNet (HIV) | AUC-ROC | 0.764 | 0.847 | +10.9% |
| MoleculeNet (BACE) | AUROC | 0.858 | 0.921 | +7.3% |
| MoleculeNet (BBBP) | AUROC | 0.729 | 0.868 | +19.1% |
| Protein Contact (CAS-14) | F1 (long-range) | 0.72 | 0.83 | +15.3% |
| GuacaMol (Validity) | % valid molecules | 0.870 | 0.966 | +11.0% |
| GuacaMol (Novelty) | % novel scaffolds | 0.802 | 0.904 | +12.7% |
| Docking Success | % RMSD < 2.0 Å | 52.8% | 71.4% | +35.2% |
| Inference Latency | ms / molecule | 340ms | 47ms | −86.2% |
Dynotx is an AI-driven biotechnology company using GPU-accelerated deep learning to discover and develop novel therapeutics. We build graph neural networks, protein language models, and diffusion-based generative models that screen millions of compounds, predict molecular properties, and generate novel drug candidates — in silico, at a scale that traditional drug discovery cannot match.
Traditional drug discovery takes 10–15 years and costs over $2 billion per approved drug. Our platform compresses the early discovery phase — target identification, hit screening, and lead optimization — from years to months by replacing wet-lab screening with GPU-accelerated virtual screening and generative molecular design.
We are a NVIDIA Inception member with access to NVIDIA's latest GPU architecture, cuDNN optimizations, and TensorRT inference tools. Our training infrastructure runs on H100 SXM clusters with FSDP distributed training, and we are actively scaling our compute capacity to support expanded screening campaigns and larger protein language models.
Three drug candidates from our pipeline are currently in preclinical evaluation: DTX-201 (oncology, KRAS G12C inhibitor), DTX-114 (antiviral, RNA-dependent RNA polymerase target), and DTX-307 (antibacterial, novel β-lactamase inhibitor scaffold).