plmMSA

A protein language model (PLM) based MSA server. Drop-in replacement for MMseqs2 MSA generation in ColabFold / Boltz / Protenix, powered by Ankh-CL + ESM-1b retrieval and ProtT5 / Ankh-Large pairwise alignment.

DeepFold-PLM is under-cited largely because obtaining plmMSAs is the bottleneck — a hosted server makes the method trivially reusable.

---

For end users

Public endpoint: https://plmmsa.deepfold.org (anonymous; no token required for MSA submission).

Submit a job in the browser

The static web UI lives at https://plmmsa.deepfold.org/ui — paste a sequence, pick an aligner / mode, click Submit. Job ids are cached in localStorage; copy the URL to share (?job=<uuid> loads that specific job in a fresh browser). No accounts, no build step on our side either — it's plain HTML / JS served by FastAPI.

One-shot curl

# Submit an MSA job.
JOB=$(curl -sS -X POST https://plmmsa.deepfold.org/v2/msa \
    -H "Content-Type: application/json" \
    -d '{"sequences": ["MKTIIAL..."]}' | jq -r .job_id)

# Poll until status is "succeeded"; result.payload holds the A3M.
while [[ "$(curl -sS https://plmmsa.deepfold.org/v2/msa/$JOB | jq -r .status)" \
          != "succeeded" ]]; do sleep 5; done

curl -sS https://plmmsa.deepfold.org/v2/msa/$JOB \
    | jq -r .result.payload > query.a3m

Every knob (aligner, mode, filter, paired, score model, VDB collection, output format) is optional — the server picks sensible defaults. Full recipe with the request-field table, error codes, and Python examples: docs/submitting-msa.md.

Drop-in MSA server for folding engines

plmMSA speaks the ColabFold MsaServer wire shape under /v2/colabfold/{plmmsa,otalign}/* — point any engine's MSA-server URL at one of those prefixes and it Just Works. Two flavors:

• /v2/colabfold/plmmsa — PLMAlign (fast, ProtT5 shard-backed).
• /v2/colabfold/otalign — OTalign (Sinkhorn + position-specific DP).

Engine	Recipe
ColabFold (colabfold_batch)	colabfold_batch --host-url https://plmmsa.deepfold.org/v2/colabfold/plmmsa input.fasta out/
Boltz (boltz predict)	boltz predict --use_msa_server --msa_server_url https://plmmsa.deepfold.org/v2/colabfold/plmmsa ...
Protenix (bytedance/Protenix)	Set the MSA-server URL in the inference config to https://plmmsa.deepfold.org/v2/colabfold/plmmsa.

Swap .../plmmsa for .../otalign to use OTalign instead. Paired MSA (multimer) support lands with the paired-MSA feature (see PLAN.md); single-chain MSAs work now.

Response format

plmMSA returns A3M by default — the same format ColabFold / AlphaFold consume directly. The server stats block (result.stats) includes hits_found, hits_pre_filter, hits_post_filter, filter_applied, and filter_threshold so you can see what the pipeline kept vs. what got dropped by the Algorithm 1 step 5 filter.

---

For operators

cp .env.example .env                      # edit for your host
cp settings.example.toml settings.toml
./bin/up.sh
curl http://localhost:8080/healthz         # bare liveness
curl http://localhost:8080/health          # aggregated readiness (all downstreams)
curl http://localhost:8080/v2/version

See docs/maintenance.md for the operator runbook (service knobs, auth tokens, wipe/reload, VDB construction) and docs/cloudflare-tunnel.md for publishing the stack under your own hostname.

---

Architecture

Nine services on one Docker bridge network: api, embedding, vdb, align, worker, cache-ops, cache-seq, cache-emb; plus an optional cloudflared under the tunnel compose profile. Full API surface + data flow are in PLAN.md.

The public surface (/v2/*) enforces input validation (1022 res/chain, 16 chains paired), per-IP + per-token rate limiting, body-size caps, idempotent submission, queue backpressure, and an audit log. See docs/submitting-msa.md §2 for the full limits + error-code table.

Performance (k=1000 CASP15 targets, current trunk):

Target	Length	PLMAlign	OTalign
T1132	102	~10 s	~22 s
T1104	106	~27 s	~28 s
T1120	235	~30 s	~42 s

PLMAlign is roughly 2× faster because ProtT5's shard store + Redis path index make the target-embedding phase disk-only. OTalign's Ankh-Large score model still re-embeds per job; a fused-kernel Sinkhorn (FlashSinkhorn) and Ankh-Large shard store are on the roadmap.

Observability

• Structured JSON logs across every service (plmmsa.access, plmmsa.access.{embedding,vdb,align}, plmmsa.audit, per-service error loggers). 5xx errors carry the full traceback; 4xx log as warnings with method / path / code.
• X-Request-ID end-to-end: stamped by api on entry, echoed on every response, threaded to embedding / vdb / align on api's internal calls, and rebound on the worker when a job is claimed so the orchestrator's downstream calls carry the same id. Grep one trace with docker compose logs ... | jq 'select(.request_id == "<rid>")'.
• Aggregated /health fans out to each downstream (HTTP + Redis ping) and reports per-service readiness; /healthz stays a bare liveness probe for compose / CF edge.
• Prometheus /metrics on api (request count, latency, in-flight gauge). Per-service metrics on embedding/vdb/align/worker is a deferred follow-up (see PLAN.md).
• Completed MSAs are cached on cache-emb keyed by a canonical submit hash — repeat submissions return immediately with result.stats.cache_hit = true. Clients can opt out with {"force_recompute": true} on POST /v2/msa.

License

MIT. See LICENSE for our copyright and NOTICE for upstream attribution and model-weight license information.