AI Infrastructure
for Drug Discovery
Run AlphaFold 100x faster on our cloud native genomics infrastructure. Fast, 10x cheaper, and extremely scalable.
AI Inference
Run your favorite models on Rafflesia
Rafflesia serves the fastest endpoints for the most popular life-science models: AlphaFold, ESMFold, ESM2, and more. Send a sequence, get a result in seconds, with the provenance that says exactly which weights produced it.
Multiple Sequence Alignment
Rafflesia offers the fastest and most accurate MSA
Building an MSA is the slowest step in most in-silico drug discovery pipelines. Rafflesia serves that search from immutable indexes in object storage, giving you accurate results in seconds, on tree-of-life-scale corpora.
Rafflesia lets you run Alphafold 100x faster.
AI agents are ushering a new era in biology, reasoning across billions of sequences to surface connections, accelerate discovery and invent new medicines.
Rafflesia is building the fastest and most scalable cloud infrastructure for the AI era.
We perform the fastest prediction for Alphafold, ESMFold and many more, with a simple, plug-and-play, agent-ready API.
If you want to deploy rafflesia on your own private cloud, contact us.
- Read query142 aa · hemoglobin subunit alpha
- Alignaffine Smith-Waterman, BLOSUM62
- Rankby alignment score
The fastest structure prediction at every scale
Best-in-class wall clock for AlphaFold, ESMFold, and every model we serve, measured against the pipelines teams run today.
Rafflesia returns an AlphaFold structure in 3 seconds, 100x faster than the reference AlphaFold2 pipeline.
Wall clock for one structure prediction: Rafflesia 3 seconds, ColabFold 60 seconds, AlphaFold2 300 seconds.
Accelerate Drug Discovery with Rafflesia
Security
Rafflesia is trusted by bioinformatics researchers, fast-growing biotech startups, pharma companies, and frontier AI labs. Our core infrastructure was built to comply with high standards of security, compliance, and privacy.
- SOC 2 Type 2
- Zero data retention available
- HIPAA compliant
- GDPR compliance
- Single sign-on (SSO/SAML)
- Single-tenant or BYOC
- Available on every cloud
- Microsoft AzurePending
Join Rafflesia
We're a small and mighty team of infrastructure engineers, statistical physicists, and bioinformaticians. We're building the foundational cloud infrastructure for life-science research.
Contact Our Team
FAQs
Rafflesia runs open biological models so you do not have to host them. Send a protein, a genome window, a structure, or a small molecule, and get back a predicted structure, an embedding, or a set of homologs.
The GPUs, the weights, and the reference databases stay on our side. AlphaFold2's search databases alone are about 2.2 TB of that.
AlphaFold does not read your sequence alone. It first gathers the known relatives of your protein into a multiple sequence alignment, a stack of related sequences that hands the model an overview of the evolutionary landscape around your query. That overview is what makes the prediction faster and more accurate, and it is not free.
Building it means searching across billions of sequences, and that search is what bottlenecks structure prediction. We have built a unique MSA infrastructure that is 100x faster, making AlphaFold inference much faster.
Folding and complexes: Boltz-2, Chai-1, Protenix v2, AlphaFold2, RoseTTAFold3, and ESMFold-2 in both its standard and fast variants. Protein embeddings: ESM Cambrian at 300M, 600M, and 6B, ESM3 1.4B Open, and ESM-2 at 8M and 35M. DNA: Evo 2 7B.
That is fourteen live model endpoints, alongside first-party homology search. The catalog carries more under a preview badge, AlphaFold2 Multimer and standalone MSA among them, and a preview endpoint is not callable until its release goes active. Everything live is reached the same way, so switching model is a change of one field rather than a change of integration.
In: sequences, structures, alignments, and small molecules. Each endpoint publishes its own schema and limits, so a malformed input is refused at the edge with a typed error instead of failing halfway through a job.
Out: the model release that ran, the structured result, URLs and hashes for any artifacts, timings, and what the run cost. Predictions write their artifacts to durable storage, and every run stays fetchable by id long after the call returns.
Yes, against public releases or a private corpus of your own. A result gives you the aligned pair, query and target coordinates, coverage, the score, and a significance estimate calibrated against the space that was actually searched. The homology page covers the engine and the benchmarks.
Yes for the deterministic primitives. Each run records its model release, parameters, seed, database snapshot, input identity, and artifact hashes, and repeating it against the same release returns the same bytes. We check that rather than assert it: five folds of one input, across staging and production, returned a single identical structure hash. Anything you ask to sample, such as design at a temperature, varies with its seed by design.
On our hosted API, as an isolated single-tenant instance, or inside your own AWS, GCP, or Azure account, where compute and storage stay in your VPC under your IAM. The contracts are identical across all three, so moving between them changes an endpoint and nothing else.
Open the dashboard and run a model against a public release, or call the HTTP API from the CLI or a generated SDK. They agree with each other because they are all generated from the same contract. Pricing is usage-based, and the catalog shows each endpoint's schema, release status, and price before you run anything.