How LynxKite Supports Drug Discovery Pipelines

Updated

LynxKite is a GPU-accelerated graph analytics platform that runs an entire drug discovery pipeline — target identification, compound screening, toxicity prediction, and benchmarking — inside a single visual workflow. It fits teams doing computational biology and drug discovery who need graph reasoning, machine learning, and chemoinformatics on datasets too large for a laptop. The platform is open core: free and open source, with a commercial tier for production teams.

The pipeline stages LynxKite covers

The site describes a four-stage flow, each stage feeding the next:

  1. Identify the target with graph reasoning — model biological relationships as a graph and reason over them to surface candidate targets.
  2. Screen compounds — molecular dynamics trajectory prediction is cited as 50,000× accelerated.
  3. Predict toxicity — property prediction runs as a stage in the same workflow.
  4. Benchmark against your protein of interest — compare candidates against a chosen protein.

Because these stages live on one canvas rather than in separate tools, outputs from one stage become inputs to the next without export/import steps.

What the live workspace actually shows

The demo pipeline (compound-screening.lk) is a working canvas where you click any box to inspect it, then run it on GPU and watch the orchestrator execute stage by stage. The visible stages:

  • Knowledge graph — source with 4.2M nodes
  • AI & agents — BioNeMo, NIM, Nemotron 3.5 Agent, GNN Model
  • Model retrain — a loop that reinforces and retrains the GNN
  • Chem & output — RDKit filter (Tanimoto ≥ 0.7), then visualize
  • Output — ranked candidates

GPU nodes spin up and down automatically as the pipeline runs.

Graph reasoning plus ML plus chemoinformatics in one place

The platform combines three capabilities that are usually separate:

  • Graph operations — 600+ available, with RAPIDS/cuGraph under the hood for graphs that don't fit on a laptop.
  • Machine learning — GNN models with retraining loops built into the workflow.
  • Chemoinformatics — 2D/3D molecule visualization and RDKit integration, described as built in rather than bolted on.

Integrations with NVIDIA BioNeMo and RDKit support GPU-accelerated, AI-guided molecule generation and screening.

LLM agents orchestrating models from genes to molecules

LynxKite is described as agentic, multi-scale modeling: LLM agents orchestrate open and proprietary models across scales, from genes to molecules. A practical consequence for budget planning: open-model GPU microservices start and stop with the workflow, so idle GPU time does not accumulate on the programme budget.

Running your own models

Any Hugging Face model can run as a LynxKite Inference Microservice. A LIM turns a plain Python row function into a scaling GPU service on Kubernetes with a single @lim decorator — the site's framing is that "biology moves faster than packaging," so wrapping a model shouldn't require a separate deployment project.

Production use and hardware foundation

LynxKite powers Biophytis' AI-driven longevity drug-discovery platform, the MASSIVE project, showcased with Biophytis, Nebius, and Accenture at NVIDIA GTC 2026. The company is an NVIDIA Inception Premier member and builds on NVIDIA RAPIDS and CUDA.

When this fits your work

Choose LynxKite if you need graph reasoning and molecule analysis in one workflow, your graph data exceeds laptop memory, and you want GPU services that scale down between runs. If your pipeline is small enough for CPU-only tools, or you need a fully managed hosted service, the open-core model means you'd evaluate the commercial tier separately — the site lists a free open-source option and a commercial tier for production teams, but does not publish pricing.