PHGDH Autonomous Drug-Discovery Project — Plan v2
Last reorganized: 2026-04-18. Project target: novel allosteric/RBD-site PHGDH modulators for neurodegeneration (Alzheimer’s focus).
1. Architecture (3 tiers)
| Tier | Role | Substrate |
|---|---|---|
| 0 | OS-level daemons (no LLM) | LaunchDaemons + cron + scheduler_loop.sh |
| 1 | Autonomous agents | jakeclaw (Gemma 4 26B @ L0), qwenclaw (Qwen 3.5 @ W0), Hermes |
| 2 | Human consultant | Claude — design + unstick only, minimize cost |
Hosts: L0 = 192.168.68.200 (inference). W0 = 192.168.68.201 (agents, dashboard, scheduler). Cheaha = UAB HPC via SSH alias cheaha (user=jakechen).
2. 5 Aims
| Aim | Title | Status | Completed tasks |
|---|---|---|---|
| 1 | Automated data pipeline + enrichment | 🟡 in progress | 1.1 |
| 2 | Computational analysis of PHGDH inhibitors | ✅ complete | 2.1, 2.2, 2.3 |
| 3 | Literature synthesis & knowledge graph | 🟡 in progress | 3.1 |
| 4 | Structure-based virtual screening on Cheaha | 🟡 active | 4.1 |
| 5 | Scientific deliverables (paper, slides, wiki) | ⚪ pending | — |
| ∞ | Infrastructure (cross-cutting) | ✅ complete | liveness, notifier, AGE, resilience, auto-resolver, plan-sync |
3. PLEASER iteration loop (per task)
7 stages, single-char codes: P L E A S X R (X = eXpense; see STAGE_CODES.md).
One active task at a time. AGE scored 1–9 reverse-NIH. Stage-timeout thresholds enforced by liveness → auto-resolver.
4. Infrastructure skills built (2026-04-18)
| Skill | Purpose |
|---|---|
liveness-audit | 10-channel health monitor every 5 min |
age-scoring | Accuracy + Growth + Expense per iteration vs 7-day baseline |
notifier | Telegram alerts on milestones (dedup + rate-limit + queue) |
resilience | Atomic writes, retry queues, hourly checkpoint, boot recovery, scheduler_loop |
auto-resolver | Self-healing: per-channel playbook + 4-step escalation ladder |
plan-sync | (new) Auto-updates BACKLOG completion status from COMPLETED + artifacts |
5. Data assets (on W0 ~/.openclaw/workspace/data/)
aim2_top20_inhibitors.csv— 20 rows, pv ≥ 7, rankedaim2_pdb_binding_sites.json— 5 Å residue shells per PDB ligandaim3_structures/— 4NJN, 6RIH, 6PLF, 7S3R, 2G76 (PDB files)aim3_pubmed_phgdh_allosteric.json— 29 papers 2024–2026aim4_smiles/smiles_potency_sorted.tsv— 1011 moleculesaim4_docking_results/— Vina dock poses (currently 1; scaling to 20)
6. Key decisions & conventions
- Publish gate: PUBLISH-01 — Jake triggers publish, daemons never push
- Stage codes: never rename
X(eXpense) - Scheduling: macOS SSH can’t install LaunchAgents/cron → scheduler_loop nohup with 3 resurrection paths
- IPs: L0=.200, W0=.201 (NOT .133/.134)
- Cheaha: SSH alias
cheahaas userjakechen, keyid_ed25519, no VPN for login node
7. Next 48 h
- Task 4.2 — scale Vina to top-20 inhibitors × 6RIH K4T pocket
- Task 1.2 — Hermes data→knowledge enrichment loop
- Task 5.2 — Repo README + pipeline diagram
- Move L2 report cron 17:00 → 23:00 (needs user-terminal sudo)
- User-gated publish of 3 wiki drafts + 20 concept-hub stubs
8. Automatic plan maintenance
Starting 2026-04-18, plan-sync runs:
- Every hour via scheduler_loop — scans COMPLETED.md + artifact files, auto-updates BACKLOG.md task statuses (pending → DONE when evidence found)
- At every R-stage exit — post_r_watchdog calls plan-sync before promoting
- Writes
project-state/PLAN_STATUS.md— a one-page human-readable dashboard - Telegram 📋 notice when any status flip occurs
No more stale plans.
Comments