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.

Scope

Run the Dayhoff reference app locally to generate novel protein sequences, score variants, scaffold motifs, and (for some variants) condition on homologous sequences.
Three ways to run the reference app, in increasing order of capability and setup cost:
1. Path A — Cached demo: frontend only (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.
2. Path B — Proxy mode: Flask proxy on :8000 + frontend on :5173, calling a dayhoff-multi Azure ML endpoint. Two terminals, no local model load. Live arbitrary-prompt generation.
3. Path C — Fully local: adds a self-hosted score server on :5001 (and optionally a fold server) running on a CUDA-capable NVIDIA GPU. Three to four terminals.
Hosted models are also published to Hugging Face and Azure AI Foundry — use the proxy in backend/app.py to call a dayhoff-multi AML endpoint, or use the transformers library directly against the Dayhoff model collection.
Four model variants (3b-GR-HM-c, 3b-GR-HM, 3b-UR90, 170m-UR50-BRn) with different strengths — see docs/about-dayhoff.md for the picker.
Bundled biosecurity guardrail: 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.

Prerequisites

Windows users: Dayhoff prototype is developed and tested on Linux / WSL2. On Windows 11, run everything inside a WSL2 distro (Ubuntu recommended) with NVIDIA drivers installed on the Windows host (not inside WSL). Native Windows Python / PowerShell is not supported. See Microsoft's WSL install guide and NVIDIA's CUDA-on-WSL guide. (Path A — cached demo — is the one exception: if a Windows user just wants to see Dayhoff outputs in the UI, npm run dev from the frontend on native Windows works fine because it never invokes the Python backend or the GPU.)
Linux (or WSL2) with a CUDA-capable NVIDIA GPU for Path C. Path B needs only network access to the AML endpoint. Path A needs only Node.js. CPU works only for the 170M variant on toy sequences and is unusably slow for the 3B variants.
Python 3.10+ for the backend (Path B and Path C); Node.js 18+ for the frontend (all paths).
Disk space:
- Path A (cached demo): ~200 MB for node_modules.
- Path B (proxy-only, calling a remote AML endpoint): ~1 GB for Python deps.
- Path C with the 170M variant: ~2 GB additional for weights.
- Path C with all four variants loaded: ~20 GB GPU memory and ~25 GB disk for weights.

Workflow

Load docs/about-dayhoff.md when users ask what Dayhoff is, how it works, which variant to use, or for benchmark numbers.
Follow docs/quick-start.md to get the reference app running locally — proxy first, then frontend, then optional score / fold servers.
For code examples (generation, fitness scoring, motif preservation, batch sampling, FASTA / CSV / JSON export), route to docs/application-patterns.md.
For loading the Dayhoff Atlas datasets (GigaRef, BackboneRef, DayhoffRef), route to docs/data-integration.md.
For extending the prototype (new endpoints, custom scoring filters, batch jobs, database integration), route to docs/prototype-expansion.md.
For GPU sizing, throughput tuning, and scaling, route to docs/performance-guide.md.
For errors, route to docs/troubleshooting.md.
Before any sequence is exported, shared, or sent to a wet lab, route to docs/responsible-use.md.

Operating rules

Windows users: Dayhoff is Linux/WSL2-only for Path B and Path C. On Windows, all backend / score-server commands must run inside a WSL2 distro with NVIDIA drivers installed on the Windows host. Native Windows Python and PowerShell are not supported. Point Windows users at Microsoft's WSL install guide and NVIDIA's CUDA-on-WSL guide before running any Path B/C step. (Path A — cached demo — is the one exception: pure-frontend npm run dev works on native Windows because it never invokes Python or the GPU.)
No-GPU demo path exists: If the user has no GPU, no AML endpoint, or just wants a quick look at what Dayhoff produces, route to Path A in 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."
GPU probe first (Path C only): Before suggesting Path C (local model loading), run 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.
First-run weight download (Path C): First call to the local score server downloads model weights from Hugging Face. The 170M variant is ~1 GB; loading all four variants is ~25 GB. Tell the user this will be slow and must not be interrupted. Subsequent runs reuse the cache. Recommend huggingface-cli login and HUGGINGFACE_HUB_ENABLE_HF_TRANSFER=1 to avoid throttling.
Multi-terminal discipline: Path A needs one terminal (frontend on :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.
Local credentials handling: Copy .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.
Honest scope: The reference app is a research prototype. Generated sequences are model predictions of plausibility, not validated designs. The bundled 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.
Biosecurity gate: Before assisting with any sequence export, sharing, or wet-lab handoff, surface 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.
Variant selection matters: When a user asks for sequence generation, pick the right variant (see 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.
Execute, don't display: When terminal execution is available, run quick-start commands directly rather than printing bash blocks for the user to copy.
Always offer the next step: After loading any explainer-style doc (e.g. 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.

Routing

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

Learning Path

docs/about-dayhoff.md — Understand what Dayhoff is, the four model variants, and which one fits your task
docs/quick-start.md — Run the reference app locally and generate your first sequences
docs/responsible-use.md — Read the biosecurity guardrails and pre-export checklist before exporting anything
docs/application-patterns.md — Build your own generation, scoring, and export workflows
docs/data-integration.md and docs/prototype-expansion.md — Scale up with Atlas datasets and custom prototype features

Reference Links

Research paper (preprint): https://aka.ms/dayhoff/preprint
GitHub: https://github.com/microsoft/dayhoff
Hugging Face model collection: https://huggingface.co/collections/microsoft/dayhoff-atlas-6866d679465a2685b06ee969
Azure AI Foundry model card: https://ai.azure.com/catalog/models/Dayhoff-170m-GR

Assets

assets/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.