A working IL-17A peptide selection + simulation plan, and the exact private OpenRouter setup — installed by your agent, not by you. Works with opencode, omp, and the DeepSeek harness, on this machine or your always-on Tailscale box.
Stefano Pineda · 29 October 2026 · ~20 min read · for opencode / omp
2.6 nMHAP SPR Kd
4,401 ŲIL-17A / IL-17RA BSA
1,940 ŲHAP / IL-17A BSA
926specific hits in the demo
19tests passing
BRIEF · CROHN’S DISEASE · ORAL PEPTIDE
Same loop, harder modality: an oral peptide for Crohn’s disease.
The IL-17A plan below is the template. Retarget it at a Crohn’s cytokine and the binding problem stays familiar; the constraint that decides the program becomes oral survival. Here is what the model carries, in order — and where it stops.
Anti-TNF (infliximab, adalimumab, certolizumab pegol), anti-p40 (ustekinumab) and anti-p19 (risankizumab) all work. They are ~150 kDa proteins given IV or SC, systemically immunosuppressive, immunogenic in a fraction of patients, and expensive to manufacture. A 10–20-mer that occupies the same functional epitope could be oral, gut-restricted, cheap to synthesise and minimally immunogenic. That is the hypothesis. The model helps with every dry-lab step before the bench; it does not decide the biology.
OPEN ROUTE
Fewer false-positive refusals
A frontier closed model routes every biology prompt through a vendor classifier before you see an answer. Ordinary, publishable immunology — “list the hotspot residues in the TNF-α / adalimumab interface”, “design a 12-mer that competes with risankizumab” — sits close enough to the patterns those classifiers block that false-positive refusals become a tax on the loop.
deepseek/deepseek-v4.1-flash-1017 on a ZDR route changes three things: the policy is inspectable and stable; the same model runs locally on vLLM when a candidate sequence cannot leave the lab (§02); and the 1,048,576-token context holds the papers, the PDB files, and the sort-seq table in one session. Not an oracle, and not a replacement for dual-use review — a benign, marketed-cytokine program just should not need a vendor’s permission to be discussed.
Four decisions, and what the model does with each
Decision
The real question
What the model produces
Target
TNF, p40, or p19 — which epitope both blocks signalling and is reachable by a peptide?
Fab footprints from the co-crystals (3WD5, 4G3Y, 3HMX, 3D85), receptor-site overlap, pocket and hotspot ranking.
“Same site”
A Fab paratope is a flat, discontinuous surface built from six CDR loops. A 10–20-mer presents one continuous face. Can a peptide occupy the functional hotspot?
Re-frames the goal from paratope mimicry to competitive occupancy, and reports the residual mismatch instead of hiding it.
Modality
Linear, cyclic, stapled, N-methylated, D-amino acid — which survives the gut and still binds?
Protease-site and liability scan, cyclisation and stapling options, synthesis route and cost, SAR positions for the library.
Oral
Proteolysis and permeability. Is gut-restricted action enough, given the target sits in the intestinal mucosa?
Assay design — simulated gastric and intestinal fluid, Caco-2, a gut-restricted rationale — not an invented PK claim.
STRUCTURES → NUMBERS
Epitope math on a real complex
Load the Fab–cytokine co-crystal, split the chains, compute interface residues, buried area, and the overlap with the receptor site. The same Kabsch / SASA / BSA code already on this page runs on a Fab complex unchanged.
LIBRARY → TRIAGE
Focused, not 10⁹ random
Mutate the epitope-facing positions; choose display (phage, yeast, mRNA); read sort-seq every round; predict the complex against the Fab-bound template; score the interface; spend MD and FEP on the top 50.
ASSAY → DECISION
Compete with the drug, not just bind
Competition ELISA against the antibody (epitope binning), SPR/BLI, a TNF-driven NF-κB or IL-23-driven STAT3 reporter, and intestinal organoid/barrier models — plus the protocol and the go/no-go.
A four-week dry-lab sprint
Week
Work
Readout
1
Parse 3WD5 / 4G3Y / 3HMX / 3D85; compute Fab footprint, receptor overlap, pocket character; choose the target and the antibody to beat.
One page: residues, overlap, go/no-go
2
Liability scan and modality choice; design the focused library at epitope-facing positions; order it.
A library that samples the hotspot, not 10⁹ random peptides
3
Predict complexes on the Fab-bound template; score interfaces; MD the top 50; rank by overlap, stability, and synthesisability.
100–200 synthesizable candidates
4
SPR/BLI; competition ELISA against the antibody; cell neutralisation; simulated gastric and intestinal fluid stability.
Competes at <10× the antibody’s Kd, survives gut fluids
What this does not prove
“Same site” is not “same molecule.” A Fab paratope is a flat, discontinuous surface assembled from six CDR loops; a 10–20-mer presents one continuous face. The realistic goal is competitive occupancy of the functional hotspot, not a peptide copy of the antibody.
The approved antibodies do not share a site. Infliximab (4G3Y) and adalimumab (3WD5) occupy different faces of the TNF protomer; risankizumab (p19) and ustekinumab (p40) bind non-overlapping surfaces of IL-23. “The antibody site” is a choice — pick the one whose occupancy actually blocks signalling.
Oral is the dominant risk. Proteolysis and permeability are empirical. Nothing here predicts oral bioavailability; put SGF/SIF and Caco-2 in week 4, not month 12.
Cytokine redundancy. A single-epitope peptide is narrower than a p40 antibody, and Crohn’s is heterogeneous. Select the patient segment, not just the molecule.
IP. The antibodies are patented; the epitope usually is not, but sequence and use may be. Freedom-to-operate review before scale-up, as with HAP.
00
Let your agent do the setup
You do not run the commands. opencode (already installed), omp, or any shell-capable agent can do the whole setup — on this machine or on your always-on Tailscale box.
Paste this into opencode or omp. It installs or verifies the harnesses, puts the key in ~/.env, mirrors it to dsh, merges the configs, verifies ZDR routing, and reports what changed.
agent prompt — paste into opencode / omp
You are setting up a private DeepSeek route on this machine. Do the work yourself; do not ask me to run commands. Ask me only for the three things at the end.
Context: https://stefanopineda.com/bio-with-deepseek
Goals
1. Install or verify the harnesses.
- omp >= 18.2.6 (omp --version). If omp is missing, try its documented installer; if unavailable, configure dsh only and report.
- dsh: npm i -g @deepseek-ai/dsh if missing; dsh --version.
- Detect macOS vs Linux. Prefer user-level installs; no sudo unless required.
2. One key location: ~/.env.
- Ensure ~/.env exists and is mode 600.
- If OPENROUTER_API_KEY is absent, stop and ask me to paste it into ~/.env. Do not echo it.
- If the process environment already defines OPENROUTER_API_KEY, use it and report; do not duplicate.
- If a higher-precedence .env already defines OPENROUTER_API_KEY (cwd/.env, ~/.omp/agent/.env, ~/.omp/.env), report that and use it; do not duplicate.
- Mirror the value to ~/.dsh/.env as OPENROUTER_DSH_KEY (mode 600). Never print either value; verify only that each is non-empty.
3. omp config.
- Back up ~/.omp/agent/models.yml and ~/.omp/agent/config.yml.
- Add or merge under providers.openrouter.modelOverrides:
deepseek/deepseek-v4.1-flash-1017:
compat:
openRouterRouting:
data_collection: deny
zdr: true
allow_fallbacks: true
- If models.yml begins with the model-sync "Generated by model-sync" header, do not hand-edit it. Put the same override in the model-sync registry or a dedicated omp profile; report which path you used.
- In config.yml set defaultModel and every text role (default, smol, slow, plan, tiny, task) to openrouter/deepseek/deepseek-v4.1-flash-1017.
4. dsh config.
- Back up ~/.dsh/settings.yaml.
- Merge:
llm-pi-ai:
providers:
openrouter:
apiKeyEnv: OPENROUTER_DSH_KEY
modelOverrides:
deepseek/deepseek-v4.1-flash-1017:
compat:
openRouterRouting:
data_collection: deny
zdr: true
allow_fallbacks: true
agent-default-model:
provider: openrouter
model: deepseek/deepseek-v4.1-flash-1017
- Do not remove existing local providers or credentials.
5. Privacy.
- Verify https://openrouter.ai/api/v1/endpoints/zdr returns endpoints for deepseek/deepseek-v4.1-flash-1017 and that the first-party DeepSeek provider is absent. Report the count and 3-5 provider names.
- If you have browser control, open https://openrouter.ai/settings/privacy and set: provider training off for paid and free; ZDR for "All other models"; OpenRouter input/output use off; private logging off. If you do not have browser control, print those exact toggles and stop for my confirmation. Do not claim they are set unless you set them.
6. Verify.
- omp models find deepseek
- omp --model openrouter/deepseek/deepseek-v4.1-flash-1017 -p "Reply with your model id and upstream provider."
- dsh --dump-config
- dsh --profile headless "Reply with your model id and upstream provider."
- Confirm both use the DeepSeek model and a ZDR-eligible provider, not first-party DeepSeek.
7. Tailscale (optional).
- If I give you a host, run tailscale status and tailscale ssh <host>; repeat steps 1-6 there.
- Keep the key in the remote's ~/.env (mode 600). Do not copy it in plaintext through logs; Tailscale SSH is fine.
- Report the remote verification separately.
8. Report.
- Files changed and backups.
- Key locations (paths only; never values).
- Privacy state and verification output.
- The one remaining human action (account toggles), if any.
Constraints
- Back up every file you touch; merge, never clobber.
- Never print, log, or commit secrets.
- Do not install anything already present; check versions first.
- If a step is ambiguous, choose the conservative path and report it.
- Do not ask me to run commands.
Ask me for
1. the OpenRouter key, if ~/.env has no OPENROUTER_API_KEY (I will paste it into ~/.env);
2. the Tailscale hostname, if I want the always-on box configured;
3. confirmation of the OpenRouter privacy toggles, if you have no browser control.
The manual sections below are reference, not homework. They are the exact files the agent merges, kept visible so you can review or hand them to a different agent. You should not need to type any of it.
01
Reference: what the agent writes
You do not need to run these. They are the exact files the agent merges, kept here for review and for agents that prefer explicit steps.
PATH A
omp — this harness
Put the key in ~/.env (omp loads it). The agent sets chmod 600.
omp --version
omp --model openrouter/deepseek/deepseek-v4.1-flash-1017
# or inside the TUI: /model
# omp models find deepseek
~/.omp/agent/.env also works; ~/.env is the simplest single location. If a higher-precedence .env already defines the key, that one wins; the agent checks and reports it. If you run model-sync, models.yml is regenerated: put the override in the model-sync registry or a separate omp --profile; the config.yml role mapping is not regenerated.
PATH B
DeepSeek Harness — dsh
Install the harness and confirm the version.
shell
npm i -g @deepseek-ai/dsh
dsh --version
dsh --profile headless "summarize the HAP paper and list the SAR positions"
dsh web
dsh reads ~/.dsh/.env reliably; the agent mirrors the key from ~/.env. If you launch dsh from $HOME, ~/.env also works.
Run dsh --dump-config to confirm the composed tree, then dsh --profile headless "..." or dsh web.
dsh is early software (0.1.2-rc). The config schema is the same pi-ai shape omp uses; dsh --dump-config is the source of truth for what your build accepts.
What just happened. Both harnesses put the openRouterRouting object on the wire as OpenRouter's provider field. data_collection: deny restricts the request to providers that do not collect user data; zdr: true restricts it to endpoints with a zero-data-retention policy. The model is still DeepSeek; the serving provider is a ZDR partner. The model id deepseek/deepseek-v4.1-flash-1017 is a 1,048,576-token, text+image snapshot.
Pin every role. A background role — smol, tiny, plan, task — pointed at another provider can send context you did not intend to send. The config above pins all of them to the same model. If the session must be private, also add the providers you do not want to disabledProviders in config.yml.
02
Keep the context private
ZDR is a routing policy, not a promise about the model. Layer it: account, request, allowlist, local. The agent verifies the request flags; if it has no browser control, the account toggles are the one human checkpoint.
Layer
Setting
What it enforces
Account
OpenRouter → Settings → Privacy
Opt out of provider training; enable ZDR for “All other models” (and the groups you use); keep OpenRouter input/output use off; keep private logging off.
Request
data_collection: deny + zdr: true
Only providers that do not collect data; only endpoints that do not retain prompts.
Hard pin
only: [...] + allow_fallbacks: false
An explicit provider allowlist. The request fails rather than falling back outside it.
Local
omp / dsh → Ollama or vLLM
No third party at all. The only no-third-party path.
The account settings
These are the one step the agent may hand back to you. If it has browser control, it can set them; otherwise it prints them and waits.
Open OpenRouter → Settings → Privacy. Turn off “allow providers that may train on your data” for paid models, and for free models if you use them.
Enable ZDR for “All other models” (and any model groups you use). This is the account-wide backstop; the per-request zdr: true is the enforcement.
In Observability, keep Private Input & Output Logging off. In Privacy, keep OpenRouter Use of Inputs/Outputs off. Both are off by default; the second pays a 1% discount if you turn it on. Do not, for unpublished work.
Optional: a hard provider pin
data_collection: deny and zdr: true already filter to eligible providers. If you also want to pin the exact set — and fail rather than route elsewhere — add only and set allow_fallbacks: false. The list below is a starting point; verify it, because provider policies and availability change.
hard pin — optional
# Optional: fail rather than route outside your allowlist.
providers:
openrouter:
modelOverrides:
deepseek/deepseek-v4.1-flash-1017:
compat:
openRouterRouting:
only: [deepinfra, together, fireworks, nebius, parasail, baseten]
allow_fallbacks: false
data_collection: deny
zdr: true
Verify it
shell
# ZDR endpoints for this model (the list changes; check it)
curl -s https://openrouter.ai/api/v1/endpoints/zdr \
| jq -r '.[] | select(.model_id=="deepseek/deepseek-v4.1-flash-1017") | .provider_name' \
| sort -u
# Which upstream provider actually served a request
curl -s "https://openrouter.ai/api/v1/generation?id=$GENERATION_ID" \
-H "Authorization: Bearer $OPENROUTER_API_KEY" | jq '.provider_name'
The ZDR endpoint list is live at /api/v1/endpoints/zdr. At the time of writing it returns 71 endpoints for deepseek/deepseek-v4.1-flash-1017 — and the first-party DeepSeek endpoint is not among them. If the generation record says provider_name: "DeepSeek", ZDR is not in effect; the request went to first-party.
What ZDR does not cover. The provider still processes the prompt in memory to answer it; ZDR means it does not retain it. OpenRouter stores metadata (tokens, latency), not prompt content. CSAM screening and legal preservation can retain flagged content even on ZDR traffic. Tools and plugins you enable — web search, browser, sandboxes — have their own policies; ZDR does not cover them. For patient-identifiable or truly unpublished data, use the local path.
The strongest path: local
If the context cannot leave the lab, run the same harness against a local OpenAI-compatible server. No router, no provider policy, no third party.
Run omp --model local-vllm/deepseek-v4, using the model id your server reports at /v1/models. The same route shape works under dsh's llm-pi-ai.providers; point baseURL at vLLM and set agent-default-model.
03
The unlock: run more loops before the bench
The bottleneck is not knowing what to try. It is the number of design-build-test-learn loops one researcher can run.
THROUGHPUT
Dry-lab 80%, continuously
Literature, structure parsing, library design, NGS analysis, and triage are the unglamorous 80%. A tool-using model can run them while the bench is busy.
COST
Tens of dollars, not a grant
DeepSeek V4.1 Flash is ~$0.03/M input and ~$1.14/M output on OpenRouter. A campaign’s reading and triage is a compute line, not a budget decision.
PRIVACY
Unpublished context is usable
ZDR routing changes the default for sequences, negative results, and target structures that a lab would not send to a training provider.
The worked example is IL-17A peptide selection. It is concrete enough to show the loop: a real target, a known peptide, a focused library, NGS, structure-guided triage, and a bench validation plan. The model is not an oracle; it is capacity.
IL-17 is a cytokine (not a plant cytokinin). The plan below assumes IL-17A; IL-17F or the IL-17A/F heterodimer change the selectivity panel, not the architecture.
The prior art that resets the plan
A 14/15-residue linear peptide already binds IL-17A and blocks IL-17RA: HAP, IHVTIPADLWDWIN[K], from the inhibitor complex PDB 5HHV / 5HHX (Liu et al., Sci Rep 2016). It is the seed, the positive control, and the calibration spike-in — not a reason to screen 10⁹ random peptides.
HAP benchmark
Value
SPR Kd, human IL-17A
2.6 ± 0.5 nM
SPR Kd, human IL-17A/F
4.6 ± 0.7 nM
SPR / FRET IL-17RA competition
0.9 / 0.8 nM
BJ fibroblast cell IC50
17 nM
PEG21 dimer, position 14
0.1 nM (avidity)
IL-17F binding
none up to 100 nM
Two HAP molecules bind one IL-17A dimer. The N-terminal IHVTI β-strand inserts between β0/β4; the C-terminal ADLWDWIN α-helix directly occupies the IL-17RA site. W12 inserts into a pocket made by F110/Y62/P59/R101; D8 contacts Y62/K114. The W12 pocket is not present in apo IL-17A; the holo structure is the template.
What the current selection gets wrong
What you do now
What it actually selects
Fix
Sort on raw binding / luminescence
Avidity, high-copy display, sticky clones
Two-colour binding ÷ expression gate
One sort, final pool only
Noise and founder effects
NGS every round; ≥2 replicates
No counter-screen
Streptavidin / tag / plastic binders
IL-17F, streptavidin, irrelevant protein
No off-rate
Fast-koff clones look fine at equilibrium
1–24 h chase with unlabelled IL-17A
No functional gate
Binders that do not block signalling
IL-17RA-Fc competition sort
Random biotinylation
Labels K114 and the receptor site
AviTag/BirA; SEC-MALS QC
Yeast “Kd” as affinity
10–100× avidity inflation
Synthesize; SPR/BLI
The corrected loop
01Target
AviTag IL-17A homodimer; BirA; SEC-MALS; confirm HAP at ~2.6 nM.
Display orientation, N-acetylation, and the luminescence selector
HAP’s N-terminal IHVTI β-strand is essential (Δ1–3 is inactive). Standard Aga2p display leaves the peptide N-terminus free — correct — but exposes it to proteases; the paper shows the uncapped peptide is degraded. Use a protease-deficient strain, 4 °C, EDTA-free buffers, an N-terminal integrity tag, and confirm the synthetic N-acetyl peptide. Test both orientations as controls.
The luminescence selector is excellent for an orthogonal functional readout (NanoLuc/BRET IL-17RA competition) but it is not a clean per-cell binding channel: luminescence is enzymatic and cell-state dependent. Use fluorescence for the sort if the instrument has it; otherwise sort on a luminescence ratio with an internal standard, and confirm on the plate reader.
04
The simulation pipeline
Structure is triage, not proof. Predict the complex against the holo template, score the interface, then spend the expensive physics on the top few.
Boltz-2’s affinity head is small-molecule only. It is not a peptide Kd predictor. Use the structure and confidence metrics; use Rosetta/MD/FEP for relative energetics; use SPR/BLI and a cell assay for the truth.
The template matters more than the model. HAP binding requires a conformational change that creates the W12 pocket; sequence-only prediction may not induce it. Supply 5HHV/5HHX as a template and model the homodimer (chains A/B) plus the peptide (chain C), not a monomer.
05
What was built and verified
Real coordinates, real interface math, a working sort-seq analysis, and an end-to-end demo that recovers the HAP consensus.
HAP epitope overlaps the receptor footprint at 4/49 residues; superposing 5HHV onto 4HSA (3.01 Å RMSD, 99 Cα) puts HAP residues 5–14 into IL-17RA’s space: 57 atom pairs < 2 Å, minimum 0.36 Å.
Poly-Ala ablation: BSA 1,940 → 1,609 Ų; contact density 19.9 → 15.9. The score is sensitive to side chains, not just backbone.
Sort-seq demo
20,000-variant focused library at the SAR-tolerant positions; three rounds; a counter-screen.
926 specific hits; sticky clones removed by the counter-screen.
Consensus of the top 50: IHVTIPADLWDWIN — HAP.
Precision@50 is 0.26, deliberately reported as a noise-floor diagnostic, not a performance claim: one shallow sort is noisy. Replicates and deep input sequencing are what make it quantitative.
The demo is a test of the analysis, not a binding prediction. The real experiment uses replicate sorts, deep input sequencing, and SPR/BLI on the synthetic peptide.
06
Thirty days
A focused campaign: reagents and controls first, then the library, then the bench.
Week
Work
Readout
1
Express AviTag IL-17A; SEC-MALS; biotinylate. Display HAP and scrambled HAP; two-colour FACS. Establish IL-17RA-Fc competition and IL-17F counter-screen.
HAP Kd ~2.6 nM; orientation dependence
2
Order the focused library (positions 2, 4, 5, 7, 14) with 0.1% HAP spike-in; transform; QC by NGS; MACS round 1; FACS round 2; NGS.
Coverage; HAP spike-in enrichment
3
Off-rate and receptor-competition sorts; NGS every round; retrain the sequence model; pick 100–200 candidates.
Replicate concordance; consensus
4
Synthesize the top 50–200; SPR/BLI; IL-17RA competition; cell IC50; IL-17F specificity.
Go/no-go
Go/no-go: Kd < 10 nM, competition < 100 nM, >10× IL-17F selectivity, cell IC50 < 100 nM. If the monomer stalls, the paper’s PEG-dimer route is the fastest known path to sub-nM potency; if the peptide is the wrong modality, switch to a miniprotein scaffold against the same region I.
07
What this does not prove
IP / FTO. HAP and its analogues are from a Pfizer paper/patent family. Research use is one thing; therapeutic development needs a freedom-to-operate review. Public sequence is not a licence.
Yeast avidity is not affinity. The surface is multivalent and IL-17A is a dimer; a 2:1 interaction can look 10–100× tighter than the soluble monomer. SPR/BLI on the synthetic peptide is the truth.
N-terminal stability. HAP is N-acetylated in the paper; the uncapped peptide is degraded. Display leaves the N-terminus free. Use a protease-deficient strain, 4 °C, EDTA-free buffers, and an N-terminal integrity tag.
The 3 Å superposition. The blocking clash analysis is a hypothesis. The C-terminal helix is ordered by the Fab in 5HHV; in solution it may be flexible. MD and a ternary complex are the confirmation.
Apo vs holo. The W12 pocket is absent in apo IL-17A. Sequence-only structure prediction may miss it. Use 5HHV/5HHX as a template.
Boltz-2 affinity is not for peptides. Its affinity head is trained on small molecules; do not report its output as a peptide Kd.
Simulation is triage. The demo’s precision/recall are a noise-floor diagnostic. Replicates, deep sequencing, and SPR are what make the ranking real.
Privacy is layered, not absolute. ZDR stops retention by the inference provider; it does not anonymize the prompt, cover tool backends, or override legal preservation. For the strongest guarantee, run locally.