# GPU-Required Tasks: Scope & Cost Review

**Question addressed:** Which tasks across this project were stalled or postponed specifically because no GPU was available, and what would each cost to run if GPU access is added?

**Method:** Audited all five substantive work sessions in the project (transcripts + saved artifacts). Every session that touched protein design was searched for GPU/model/deferral language, and the deferred items were cross-checked against the executable protocol specs the agent wrote *in lieu of* running them (`binder_validation_protocol.md`) and the design rosters (`binder_design_targets.csv`, `celiac_design_spec.json`). Costs are modeled at Modal on-demand rates (the compute provider already wired to this project).

---

## Headline

The entire deferred GPU backlog is **7 tasks, ~25–54 GPU-hours, ~$55–$121** at Modal on-demand rates. This is a rounding error relative to the scientific value gated behind it (in-silico validation of 5 designed therapeutic leads across two diseases). **The blocker was never cost — it was that no GPU host was configured at all.** Even generous re-run/iteration allowances keep the whole backlog under ~$150.

The single clean-pass cost (no iteration) is only **$24–$52**. The ranges below add an iteration factor (2–2.5× on generative steps) because real design campaigns re-run with tuned parameters.

---

## Where the stalls came from

Two sessions produced GPU-deferred work; the other three (dataset mining, skill publishing, methods paper) produced **no** GPU-blocked tasks — their deferrals were CPU-only or infrastructure (a macOS Command-Line-Tools boot failure, an IEDB/netMHC server outage), none of which a GPU fixes.

| Session | GPU-blocked? | Note |
|---|---|---|
| Phase 1 Dataset Analysis | Design sub-thread only | Bulk/scRNA omics all ran on CPU; only the downstream binder-design pipeline was deferred |
| Methods Paper | Documents the deferral | Explicitly ships Steps 5–6 as *specifications* because "no GPU host was available" |
| Celiac Food-Antigen Model | Yes | 96 CPU designs completed; fold-validation + co-fold + affinity maturation deferred |
| Publish Omics Skill | No | Blocked on CLT/GitHub, not GPU |
| User Uploads | No | — |

The agent's own words, repeatedly: *"No remote compute is configured, and locally I have no GPU (8 GB RAM). RFdiffusion backbone generation and Boltz-2/Chai-1 validation realistically need a GPU."* Where it could, it substituted a CPU-feasible path (ProteinMPNN/SolubleMPNN scaffold redesign) and wrote the GPU pipeline as an executable spec to run later.

---

## The exhaustive task list

Costs: **realistic** = single-pass GPU-hours × iteration factor × Modal $/hr. GPU column is the recommended card class.

| ID | Campaign | Task | Tool | GPU | GPU-hr (realistic) | Cost (USD) |
|----|----------|------|------|-----|-------------------|-----------|
| **E1** | EoE 3 leads | De novo binder backbone generation (SIGLEC6, IL1RL1, CCL26) | RFdiffusion / BindCraft | A100-40 | 10–20 | **$21–42** |
| **E2** | EoE 3 leads | ESMFold2-Fast complex triage (thousands of designs) | ESMFold2-Fast | L40S | 1–2 | **$2–4** |
| **E3** | EoE 3 leads | Chai-1 / AF-Multimer final complex ranking (ipTM/pAE/DockQ) | Chai-1 / AF-Multimer | A100-80 | 5–10 | **$12–25** |
| **E4** | EoE 3 leads | Designed-binder immunogenicity / ADA screen | T-cell epitope pred (mostly CPU) | L4 | 0.3–0.8 | **$0–1** |
| **C1** | Celiac 2 blockers | Fold-validate 96 completed CPU designs (pLDDT) | ESMFold2 / Boltz-2 | A100-80 | 2–4 | **$5–10** |
| **C2** | Celiac 2 blockers | Complex co-fold (ipTM): MICA:NKG2D & IL-15 binder | Boltz-2 / Chai-1 | A100-80 | 2.5–7.5 | **$6–19** |
| **C3** | Celiac 2 blockers | RFdiffusion affinity-maturation round | RFdiffusion + ProteinMPNN | A100-40 | 4–10 | **$8–21** |

**Campaign subtotals (realistic):** EoE 3-target leads **$36–72** · Celiac 2-target blockers **$20–50**
**Whole backlog:** **$55–121** (≈25–54 GPU-hours) realistic · **$24–52** single clean pass

---

## Dependencies & sequencing (matters more than the dollar cost)

The tasks are **not** independent — there is a hard chain inside the EoE campaign:

```
E1 (generate backbones) ──▶ E2 (triage fold) ──▶ E3 (rank complexes) ──▶ E4 (immunogenicity)
```

E4's inputs (designed binder sequences) **do not exist yet** — they are produced by E1. So the EoE campaign must run front-to-back; you cannot cherry-pick E3 without E1/E2. The Celiac tasks are more decoupled: **C1 and C2 can run immediately** (the 96 designs already exist on CPU), and C3 is an optional enhancement round.

**Cheapest high-value first move:** C1 + C2 (~$11–29) validate two already-complete celiac designs (MICA-decoy for NKG2D, IL-15 binder) — immediate payoff, no upstream generation needed.

---

## Practical notes on the estimate

- **Rates:** Modal on-demand, approximate 2026 published $/GPU-hr — A100-40 ~$2.10, A100-80 ~$2.50, L40S ~$1.95, L4 ~$0.80. Spot/committed-use and other providers (RunPod, Lambda) run 30–60% cheaper; these are conservative upper-reference numbers.
- **Serverless billing helps here.** Modal bills per-second with scale-to-zero, so short design jobs incur no idle cost — a good fit for this bursty, hours-not-days workload.
- **What's NOT in scope (won't cost GPU):** all bulk/scRNA omics mining, the 567-gene meta-signature + LOCO robustness, HLA/epitope MHC-II prediction (ran on mhcnuggets/netMHCIIpan, CPU), ESM-2 epitope conservation (already completed on CPU). None of these are stalled on GPU.
- **One-time setup:** first-run image builds (RFdiffusion, Boltz-2, Chai-1 containers + weight download) add a few GPU-minutes and a one-time cold start; negligible on this budget but real on wall-clock (~10–20 min first launch each).
- **Skill gap, not just compute:** the transcripts note **no RFdiffusion skill exists in the catalog** — E1/C3 need that tool installed/wrapped, not only a GPU. ProteinMPNN, AlphaFold2, and ESMFold skills are already present.

---

## Bottom line for the investment decision

Adding a GPU host unblocks **7 concrete, fully-specified tasks** for well under **$150** of compute — the protocols, targets, epitope hotspots, and acceptance thresholds are already written and saved. The decision is effectively free on cost grounds; the real gating items are (1) wiring a Modal (or other) GPU host, and (2) packaging RFdiffusion, which isn't yet a catalog skill.