# Native scientific worker integration Run workers in pinned native environments. Install PepPA without replacing their scientific dependencies. Registered commands receive request and result paths. Save persistent artifacts outside the temporary request directory. Return `schemas/ToolResult.json`, with a sidecar recording command, revision, seed, hashes, runtime, GPU allocation, and score conventions. ## moPPIt / PepDFM The inspected revision uses `moppit.py`; older examples mention `moo.py`. The supplied bridge uses the actual entrypoint, sets Python/NumPy/Torch seeds, uses `n_samples=1`, and obtains batches through `n_batches`. Objective names and weights follow native scorer initialization order. The entrypoint fixes step size to 1/100, so the bridge requires 100 steps. It exports canonical linear proposals. Modified and cyclic candidates need a compatible worker. Save the conversion from one-based canonical motif positions to native indices. Native CSV scores require endpoint calibration before conservative ranking. ## PeptiVerse The inspected API is `PeptiVersePredictor(manifest_path, classifier_weight_root, device=...)`. Calls use `predict_property(prop_key, col='wt', input_str=...)` and `predict_binding_affinity(col='wt', target_seq=..., binder_str=...)`. The bridge retains native JSON outputs. SMILES endpoints require a chemistry-compatible route. Native uncertainty may be entropy, ensemble variation, or an interval. Verify the endpoint-specific definition. ## PTM-Mamba Use the native package at https://github.com/programmablebio/ptm-mamba and its declared checkpoint. The inspected example imports `PTMMamba` from `protein_lm.modeling.scripts.infer` and uses explicit tokens such as `MK`. Verify the exact vocabulary for each PTM. Save `hidden_states` with a fixed pooling rule, chemical identity, token map, checkpoint hash, and partition. Align target features to frozen ESM-2 peptide features before fitting the supplied head. PTM-specific interaction labels supervise binding prediction. ## SOAPIA / SiameseCat Supply the lab's compatible checkpoint and native two-track preprocessing. Return target and countertarget distances, beta, margin, and specificity contrast. Preserve unknown labels separately from negatives. The supplied PTM bilinear head is a separate supervised adapter, with paired chemical-state labels. ## AlloGen At the pinned revision the documented API is: ```python from models.differentiable_features import DifferentiableQTheta scorer = DifferentiableQTheta(checkpoint=checkpoint_path, device='cuda:0') scorer.load_receptor(holo_path='holo.pdb', rec_chain='A', apo_path='apo.pdb', apo_chain='A') q_holo = scorer.score('design.pdb', binder_chain='B', state='holo') q_apo = scorer.score('design.pdb', binder_chain='B', state='apo') ``` Run with the native `code` directory on the Python path. Retain both state scores and their difference. Check input-feature support for the requested PTM chemistry. ## TD3B Load the trained Direction Oracle with native target representation and tokenizer. The inspected implementation returns probability of agonism in [0,1]. Its gate is `sigmoid(d * (p_agonist - 0.5) / temperature)`, with +1 for agonism and -1 for antagonism. Require successful trained-checkpoint loading. The source includes a neutral placeholder mode, which must be rejected during evaluation. ## AF3 `peppa.structure.af3_input` exports explicit CCD modifications and atom-level bonds using one-based construct positions. Run the native predictor separately and retain the input hash, chain-pair confidence, and all structures. The planned protocol uses three seeds and five samples per seed. Preserve the candidate identity across matched PTM inputs and ternary inputs. Report both peptide interfaces. ## EvoBind2 / EvoBind-Multimer Use native sequence-optimizing protocols. Molecular-glue sources are https://doi.org/10.64898/2026.08.21.746227 and https://github.com/patrickbryant1/EvoBind-multimer . Any changed sequence becomes a new candidate with a parent ID and fresh evaluations. Fixed-sequence AF3 validation is a distinct operation. Pin code, model parameters, templates, and candidate counts. Preserve upstream licensing. ## PepDFM, PepMDLM, and PepTune Register a generator worker under the shared CommandTool interface. PepDFM and MOG-DFM use the native sequence sampler and checkpoint from the supplied MOG-DFM study. PepMDLM uses the peptide-SMILES masked-diffusion sampler described in https://arxiv.org/abs/2412.17780. PepTune uses its native multi-objective search procedure. Freeze the exact repository, checkpoint, tokenizer, chemistry parser, decoder settings, and objective implementations before evaluation. These generator workers require native integration and preflight; they are not bundled model implementations. Each generator returns exact candidate chemistry, seed, parent ID when applicable, attempted and valid candidate counts, objective weights, native scores, elapsed GPU minutes, and source artifacts. A SMILES worker must serialize stereochemistry, monomers, terminal groups, and closure bonds into the Candidate representation. Use the same eligible generator portfolio in each controller comparison. Freeze chemistry and proposal allocation; compare generator routes in matched strata. ## Ranking and experimental acceptance Keep two explicit lists in the frozen task manifest: `score_requirements` for calibrated computational component scores and `experimental_requirements` for measured success. Pass the first list to `candidate_score`; pass the second to `joint_success`. This permits optimization of binding, properties, and interface proxies before a PTM-selectivity or ternary-cooperativity measurement exists. Preserve the lists and their hashes in artifacts and run metadata. A native property probability must retain its actual endpoint definition. ## Frozen computational workflow Connect native workers to `peppa.compiled` through registry callables with revision, run, and scientific validate functions. Use baseline `CommandTool` wrappers as process adapters. Unroll the two proposal cycles and all conditional computation before execution. Mandatory outputs include context-specific scores, lineage and applicability metadata, chemistry verification, and release diagnostics. Production adapters must enforce timeout and GPU/CPU/memory budgets. ## pepADMET, docking, and CMC Pin native pepADMET code/checkpoint and define endpoint-specific representation/domain checks. Preserve predictions independently of PeptiVerse, then compare or aggregate only matched endpoints. Record shared data ancestry. Docking contracts include preparation, protonation, search region, seeds, scoring function, and full pose outputs. CMC contracts include structure-derived formula/mass, monomer availability with dated source, route/closure/protection annotations, coupling liabilities, aggregation/chemical stability flags, and analytical specifications. These adapters require implementation and known-input preflight. ## Prospective outcome analysis Experimental assays are imported into a separate evaluation dataset keyed to the immutable candidate-release hash. Aggregate replicates with units, censoring, assay context, and raw artifact identifiers and sources. Preserve expressed uAb/duAb constructs separately from free-peptide dosing. Compute experimental success against the preregistered assay requirements after computational commitment.