dsh-aidd

AI-driven drug discovery is rarely a single-model problem. The real friction lives in the workflow: hit lists from virtual screening, hundreds of generated candidates, ADMET predictions, DTI model evaluations—all scattered across scripts, logs, and half-remembered command histories. When a question finally surfaces—how did this particular score come about?—the answer is often lost. dsh-aidd reframes AIDD research as a structured ReAct loop (hypothesis, experiment, observation, conclusion, iterate), with each step backed by explicit tool calls and state that persists across sessions, so work can be resumed on any machine without losing the thread.

On reproducibility, the plugin treats every meaningful output—hit lists, score tables, data-split files, model checkpoints, metric sheets—as a versioned artifact carrying a SHA-256 checksum, the generating command, input parameters, and an environment snapshot, plus an append-only provenance log. A set of AIDD data-discipline rules are baked into the preset persona: seeds must be fixed and recorded, scaffold and cold-start splits must be archived, homologous targets must not straddle train and test, and no improvement can be claimed until the baseline is reproduced. These guardrails address the classic trap where a good score does not equal a good molecule at the source.

The bundle also ships 17 skills covering the full AIDD spectrum—virtual screening, molecular generation, property and ADMET prediction, protein-target modeling, DTI modeling, model reproduction—alongside general research capabilities like literature connection and scientific review. It is aimed at individuals and teams doing computational chemistry, molecular design, or AI drug discovery inside DeepSeek Harness, especially in settings where collaboration, auditability, and workflow reproducibility matter.