npx skills add ...
npx skills add microsoft/vibe-kit --skill msresearch-dayhoff
npx skills add microsoft/vibe-kit --skill msresearch-dayhoff
Microsoft Research's Dayhoff — generates novel protein sequences and scores variants using hybrid Mamba+Transformer models trained on the 3.34B-sequence Dayhoff Atlas. Use when users mention Dayhoff by name, want to generate protein sequences from scratch or from a prompt, score mutations / variants / fitness without experimental data, design sequences in a protein family from homologs, scaffold sequences around a motif, or are setting up / running / troubleshooting the Dayhoff prototype locally.
npm run dev), no GPU, no backend. Returns precomputed real Dayhoff outputs for four bundled prompts (Cas9, insulin, DNA polymerase, SARS-CoV-2 spike) across all four model variants. Best for evaluating the UX or demoing on a laptop with no GPU access.:8000 + frontend on :5173, calling a dayhoff-multi Azure ML endpoint. Two terminals, no local model load. Live arbitrary-prompt generation.:5001 (and optionally a fold server) running on a CUDA-capable NVIDIA GPU. Three to four terminals.backend/app.py to call a dayhoff-multi AML endpoint, or use the transformers library directly against the Dayhoff model collection.3b-GR-HM-c, 3b-GR-HM, 3b-UR90, 170m-UR50-BRn) with different strengths — see docs/about-dayhoff.md for the picker.backend/sequence_screening.py runs a 3-layer Select Agent toxin screen on inputs and outputs. This is a guardrail, not a guarantee — see docs/responsible-use.md.npm run dev from the frontend on native Windows works fine because it never invokes the Python backend or the GPU.)node_modules.docs/about-dayhoff.md when users ask what Dayhoff is, how it works, which variant to use, or for benchmark numbers.docs/quick-start.md to get the reference app running locally — proxy first, then frontend, then optional score / fold servers.docs/application-patterns.md.docs/data-integration.md.docs/prototype-expansion.md.docs/performance-guide.md.docs/troubleshooting.md.docs/responsible-use.md.npm run dev works on native Windows because it never invokes Python or the GPU.)docs/quick-start.md — frontend only with cached real Dayhoff outputs for four bundled prompts. Don't gatekeep on hardware. The cache lives in frontend/src/demoCache.ts and is what the UI shows when the backend badge says "Backend offline · cached examples only."nvidia-smi and confirm a CUDA-capable GPU is visible. If none is present, default to Path A (cached demo) or Path B (remote AML endpoint) — do not try to load the 3B models on CPU.huggingface-cli login and HUGGINGFACE_HUB_ENABLE_HF_TRANSFER=1 to avoid throttling.:5173). Path B needs two (proxy backend on :8000 + frontend on :5173). Path C needs three (add local score server on :5001), and up to four with a local fold server. Each must stay running in its own terminal — never run other commands in a terminal hosting a live server..env.example to .env and fill in DAYHOFF_ENDPOINT / DAYHOFF_API_KEY for the AML proxy mode, or leave them pointing at http://localhost:5001/score for fully-local mode. Never ask the user to paste credentials into chat — always reference the .env file. Keep the DAYHOFF_* and ESMFOLD_* variable names exactly as written; the proxy reads those exact names.sequence_screening.py toxin filter catches a small Select Agent reference set — it is a guardrail, not a guarantee. Do not claim safety, function, or foldability for any generated sequence.docs/responsible-use.md and confirm the user has reviewed the pre-export checklist. Never assist with sequences targeting pathogens, toxins, or dual-use applications. Treat every generated sequence as a research draft requiring qualified human review.docs/about-dayhoff.md matrix). Reject homolog inputs for 3b-UR90 and 170m-UR50-BRn early — the score server will return 400 if a model that doesn't support homolog conditioning receives a homologs payload.docs/about-dayhoff.md), end your response with a concrete offer to advance the user. Default phrasing: "Want to try Dayhoff? Fastest path is just npm run dev in the bundled frontend — no GPU, no backend, four cached prompts (Cas9, insulin, DNA polymerase, SARS-CoV-2 spike) return real Dayhoff outputs instantly. From there we can stand up proxy mode against an Azure ML endpoint or fully local mode on your GPU when you want custom prompts." Adapt wording to context.| Doc | When to load |
|---|---|
docs/about-dayhoff.md | User asks what Dayhoff is, how it works, which variant to use, benchmarks, or scientific background |
docs/quick-start.md | User wants to run the reference app locally |
docs/application-patterns.md | User wants code examples for generation, fitness scoring, motif preservation, batch sampling, or export |
docs/data-integration.md | User wants to load Dayhoff Atlas datasets (GigaRef, BackboneRef, DayhoffRef) |
docs/prototype-expansion.md | User wants to extend the reference app — new endpoints, custom scoring, batch processing, database, validation |
docs/performance-guide.md | User asks about GPU sizing, throughput, batching, or scaling |
docs/troubleshooting.md | User hits an error: missing GPU, Mamba kernel error, weight download stall, port conflict, AML 401, HF rate limit |
docs/responsible-use.md | Before any sequence export, sharing, or wet-lab handoff; or when user asks about safety, biosecurity, or responsible use |
docs/about-dayhoff.md — Understand what Dayhoff is, the four model variants, and which one fits your taskdocs/quick-start.md — Run the reference app locally and generate your first sequencesdocs/responsible-use.md — Read the biosecurity guardrails and pre-export checklist before exporting anythingdocs/application-patterns.md — Build your own generation, scoring, and export workflowsdocs/data-integration.md and docs/prototype-expansion.md — Scale up with Atlas datasets and custom prototype featuresassets/dayhoff-prototype/ — Self-contained reference app: Flask proxy backend (backend/app.py, port :8000), local score server (backend/score/, port :5001), optional ESMFold structure server (backend/fold/), Vite + React + TypeScript frontend (frontend/, port :5173), Azure ML deployment YAMLs (infra/), and Dockerfile for App Service deployment.assets/paper/dayhoff-biorxiv-paper.pdf — Bundled offline copy of the bioRxiv preprint.