Back to home@cloud-pkumdl

dsh-pdb

dsh-pdb (Protein Design for Binder): agent-driven workflow orchestration for protein binder design. DAG-based plan visualization with human-in-the-loop review — unifying GUI-less protein design tools under one agentic pipeline.

Stars
2
Language
Created
Aug 21, 2026
Updated
Aug 21, 2026
GitHub repo

Introduction

dsh-pdb

Protein Design for Binder — an agent-driven workflow orchestration layer for protein binder design.

Note: pdb here stands for Protein Design for Binder — not the Protein Data Bank.

Why

Most protein design tools (RFdiffusion, ProteinMPNN, AlphaFold2, ESMFold, Rosetta...) are powerful but command-line only, fragmented, and hard to chain. Running a full binder design campaign means manually stitching together a dozen scripts, tracking intermediate files by hand, and re-planning when a branch fails — with no unified view of what's been tried.

dsh-pdb unifies these tools under a single agentic pipeline. An AI agent proposes a design plan; you review and refine it; the pipeline executes and reports back — with every step visualized as a living DAG.

Core Idea

Machine plans, human reviews, the graph is the alignment interface.

Unlike cryoSPARC (human authors the workflow, machine executes, graph is a record), dsh-pdb inverts control:

  1. Agent proposes — the agent generates an execution plan as a DAG (nodes = steps/strategies, edges = dependencies/derivations).
  2. Human reviews — you see the plan as a flowchart, then approve, prune, or graft branches before anything runs.
  3. Pipeline executes — approved nodes run; status flows back live (planned -> running -> success/failed).
  4. Re-plan on failure — failed branches trigger the agent to backtrack and propose alternatives, all traceable to the root.

This lowers the planning barrier: you don't need to know the optimal workflow up front — the agent suggests one, you correct it.

Features (planned)

  • Agent-proposed design plans — DAG generated from a natural-language goal ("design a binder against target X")
  • Interactive plan review — visualize, approve/reject branches, edit nodes before execution
  • Multi-strategy DFS exploration — parallel design strategies with automatic backtracking
  • Unified tool wrapping — RFdiffusion / ProteinMPNN / AF2 / ESMFold / Rosetta as pluggable nodes
  • Full reproducibility — every node records inputs, params, and provenance back to the root
  • Live status tracking — real-time node states with metrics (ipTM, pLDDT, etc.)

Backend Tool Licenses

dsh-pdb wraps existing design tools; it does not redistribute their code or weights. When packaging or using them, respect each tool's own license:

ToolLicenseNotes
RFdiffusionBSD-3-ClauseCode + weights both covered
RFdiffusion2BSD-3-Clause
ProteinMPNNMIT
LigandMPNNMIT
ESM / ESMFoldMIT
ColabFoldMIT
dl_binder_designMIT
AlphaFold2Apache-2.0 (code)Weights: CC BY 4.0 (commercial OK)
AlphaFold3Apache-2.0 (code)Weights: non-commercial, request from DeepMind, no redistribution
RosettaFree for academicCommercial use requires a paid RosettaCommons license

Status

Early development. Concept demo of the plan-review interface coming soon.

License

MIT — see LICENSE.