FEP Alternative Skill#

This skill predicts relative binding free energy or activity labels for newly designed small molecules using the SciMiner PBCNet 2.0 tool. It is a PBCNet-based alternative workflow for FEP-style lead optimization tasks, not an explicit molecular-dynamics FEP simulation.

The core SciMiner workflow uses:

• PBCNet 2.0 for relative binding affinity or activity-label prediction
• Gnina for generating docked 3D binding conformations when candidate or reference ligand poses are missing
• PDB database retrieval through the bundled rcsb-pdb-skill when the user has activity data but no protein-ligand crystal structure

When to use this skill#

• Predict relative binding free energy, binding affinity, or activity labels for newly designed molecules against a known target.
• The user has one or more reference molecules with measured biological activity against the same target and wants to predict newly designed molecules.
• The user provides a receptor PDB file, reference ligand conformations, labels, and candidate molecules, or asks the agent to complete missing receptor, pocket, docking, or conformation files.
• The user asks for an FEP alternative, fast relative activity prediction, PBCNet, or PBCNet 2.0.

Required PBCNet inputs#

PBCNet 2.0 requires these conceptual inputs:

• protein receptor PDB file
• SDF file containing 3D binding conformations of reference molecules
• activity label values for the reference molecules
• SDF file containing docked 3D binding conformations of molecules to be tested

If the user cannot provide all of these files, collect the missing information and supplement it through the workflows below. Do not send random or undocked candidate conformations to PBCNet; the candidate SDF must contain already-docked binding poses.

Method selection rule#

• If all PBCNet-ready files are available, invoke PBCNet 2.0 directly.
• If a protein-ligand crystal structure or receptor PDB plus original/reference ligand pose is available, use the original ligand position as the binding pocket, run Gnina to dock the molecules to be tested, and then invoke PBCNet 2.0 with the top 3 poses of each ligand ranked by centroid deviation from the reference pose.
• If measured reference activities are available but no crystal structure is available, search the PDB for structures of the same target with co-crystal ligands similar to the measured reference ligands. Keep candidates with ligand similarity higher than 30%, select the crystal with the highest ligand similarity and acceptable structure quality, extract the receptor and binding pocket from that crystal, establish reference poses for measured molecules, dock test molecules with Gnina, rank predicted poses by centroid deviation from the reference pose, keep the top 3 poses of each ligand for bioactivity prediction, and then invoke PBCNet 2.0.
• If target identity, reference ligand structures, activity labels, or test molecule structures are missing and cannot be recovered from explicit user context, ask the user for the missing information before invoking SciMiner.

Prerequisites#

1. Obtain a free SciMiner API key from https://sciminer.tech/utility.
2. Store it outside this repository at ~/.config/sciminer/credentials.json with JSON shaped as {"api_key":"your_api_key_here"}.
3. For SciMiner calls, read the API key from ~/.config/sciminer/credentials.json and send it as the X-Auth-Token header.
4. Never print, persist, or store the API key in prompts, logs, or repository files. Agents should remember only the credential file path.

If ~/.config/sciminer/credentials.json is not available or does not contain an api_key field, stop and tell the user to obtain a free SciMiner API key from https://sciminer.tech/utility and store it in that file. Do not try to complete the task by switching to other tools or services.

Authoritative tool-doc source (required)#

The published Markdown files under https://sciminer.tech/tool_api_files/ are the single source of truth for SciMiner provider_name, tool_name, allowed parameters, file-upload behavior, request encoding, and submission flow.

Use these SciMiner Markdown docs:

• PBCNet 2.0 -> PBCNet 2.0_api_doc.md
• Gnina -> Gnina_api_doc.md

For PDB database retrieval and ligand-similarity searches, read the bundled database sub-skill before querying public resources:

• rcsb-pdb-skill/SKILL.md

The agent MUST:

1. Resolve the selected SciMiner tool's Markdown file and read it before every invocation.
2. Never invent provider_name, tool_name, parameter names, enum values, upload-field names, content type, or submission flow from memory.
3. Extract and follow the selected SciMiner doc section's exact:
   • Base URL
   • API endpoint
   • Content-Type
   • Authentication header
   • Tool Name
   • Method
   • Parameter table, including required fields and enum values
   • File-upload instructions and example code
4. Choose the correct doc section if the selected doc contains multiple tool variants, such as reference-ligand input vs pocket-center input.
5. Cite the selected Markdown docs as the payload source in summaries.

If a user-provided parameter is not present in the selected Markdown doc section, ask for correction or drop it with an explanation.

Required workflow#

1. Build an input ledger before running tools. Track target identity, receptor PDB or PDB ID, reference molecule structures, reference activity labels and units, candidate molecule structures, binding pocket definition, and which files are already PBCNet-ready.
2. Determine whether the task fits the direct PBCNet path, the crystal-guided docking completion path, or the no-crystal PDB-retrieval path.
3. Read the PBCNet 2.0 Markdown doc before constructing any PBCNet payload.
4. If docking is needed, read the Gnina Markdown doc before docking any molecule and choose the doc section that matches the available pocket input.
5. If no crystal structure is available, read the bundled rcsb-pdb-skill/SKILL.md sub-skill and retrieve PDB structures for the same target using ligand similarity to the measured reference ligands. Keep structures with ligand similarity higher than 30%, prefer the highest similarity, and consider resolution, chain completeness, ligand relevance, and mutation state when breaking ties.
6. For crystal-guided workflows, use the original ligand position in the user-provided or selected crystal structure as the binding pocket.
7. Establish measured reference poses before docking test molecules. If there is only one measured reference molecule with a matching PDB co-crystal ligand, use that native ligand pose from the PDB as the reference pose. If there are multiple measured reference molecules represented across several PDB files, perform cross-docking among the candidate receptor structures and measured reference molecules; select the receptor that best preserves native or mutually consistent reference binding modes, then determine each measured molecule's reference pose in that selected receptor frame.
8. For no-crystal workflows, extract the protein receptor PDB file and bound ligand pocket from the selected PDB crystal. When extracting any ligand from a PDB file, validate and preserve chemical bond orders before writing SDF; do not allow conversion tools to turn all bonds into single bonds.
9. After extraction and reference-pose determination, dock molecules to be tested in that pocket.
10. When multiple docked poses are produced, calculate the centroid of each predicted ligand pose and the centroid of the matched reference pose in the same receptor coordinate frame. Compute centroid deviation as the Euclidean distance between those centroids, rank poses for each ligand from smallest to largest centroid deviation, and keep the top 3 poses of each ligand for PBCNet bioactivity prediction. If fewer than 3 poses are available for a ligand, keep all available poses and report the count. Preserve ligand IDs, pose ranks, and centroid-deviation values in the prepared SDF or companion metadata when possible.
11. Assemble the reference SDF so its molecule order matches the activity label order exactly. Preserve user-provided activity units and transformations; do not invent or convert labels unless the user requested the conversion or confirmed the rule.
12. Upload required file inputs exactly as described by the selected SciMiner Markdown docs and replace local paths with returned file_id values.
13. Write or run invocation code directly from the selected Markdown docs' base-information block, parameter table, file-upload instructions, and example code. Do not apply a shared invocation template or local registry abstraction in this skill.
14. Poll the task result and return the share_url in the final user-facing summary.

Missing-input completion rules#

• Missing receptor PDB but target and measured ligand are known: search PDB for same-target co-crystal structures using ligand similarity, require ligand similarity higher than 30%, then select the best supported crystal.
• Missing candidate binding conformations: run Gnina docking in the selected or user-provided pocket, calculate centroid deviation from the reference pose for every predicted pose, and use the top 3 poses of each ligand as the PBCNet candidate conformation input.
• Missing reference ligand SDF but a crystal ligand is present: extract or prepare the crystal ligand pose as a 3D SDF when that ligand corresponds to a measured reference molecule. Validate ligand chemistry during extraction: preserve bond orders from CCD or another trusted chemical template, repair aromatic/double/triple bonds when needed, and reject or flag outputs where all bonds have become single bonds. Otherwise ask the user to confirm the reference mapping.
• Multiple measured reference molecules across several PDB files: perform cross-docking to select the receptor structure and assign each measured molecule's reference pose before ranking test-molecule poses by centroid deviation.
• Missing measured activity labels: ask the user for values and units. Public database activities may be used only when the user asks for retrieval or confirms the selected measurements.
• Missing test molecules: ask for SDF, SMILES, or another structure file format accepted by the selected Gnina doc section.

File upload rules#

• Upload every required file parameter described by the selected SciMiner Markdown doc before invocation.
• Replace local paths in parameters with the returned file_id strings.
• Use the upload form field documented by the selected SciMiner Markdown doc.
• Skip optional file parameters that the user did not provide.

Expected result format#

json

{
    "status": "SUCCESS",
    "result": {...},
    "task_id": "xxx",
    "share_url": "https://sciminer.tech/share?id=<task_id>&type=API_TOOL"
}

Workflow guidance#

• Complete receptor, reference SDF, labels, and docked candidate SDF -> PBCNet 2.0.
• Receptor or crystal structure plus candidate molecules but no candidate poses -> Gnina, centroid-deviation ranking, top 3 poses per ligand, then PBCNet 2.0.
• Reference activities but no crystal structure -> PDB ligand-similarity retrieval, native-pose or cross-docking reference-pose assignment, receptor and pocket extraction with ligand bond-order validation, Gnina docking, centroid-deviation ranking, top 3 poses per ligand, then PBCNet 2.0.
• Ambiguous labels, units, molecule ordering, target identity, or reference mapping -> ask for clarification before submitting PBCNet.

Notes#

• Use the selected SciMiner Markdown docs under https://sciminer.tech/tool_api_files/ as the authoritative source for payload construction and invoke-method details.
• Read the SciMiner API key from ~/.config/sciminer/credentials.json and send it as the X-Auth-Token header. Do not print or persist the API key in prompts, logs, or repository files.
• If ~/.config/sciminer/credentials.json is missing or does not contain an api_key field, stop and tell the user to obtain a free SciMiner API key from https://sciminer.tech/utility and store it in that file.
• provider_name and tool_name must exactly match the selected Markdown doc.
• Use the selected Markdown doc to determine PBCNet label_list formatting, Gnina pocket specification, file inputs, parameter placement, and any tool-specific submission details.
• Important: when summarizing results to users, attach the share_url links of every successful task at the end so that users can view the online results, rather than showing file download links.
• For long-running tasks without a fixed ETA, poll for no more than 6000 seconds; if the task is still running, stop polling and return the current task_id and share_url so the user can check later.