aurekaresearch/opendde-harness/developability-filter

Developability Filter

Overview

Use this skill to assess antibody developability — pI, TAP, and CDR liability checks. When the check_antibody_developability tool is available it delegates to a compute backend; when the backend is not installed the agent falls back to LLM-based assessment using sequence properties.

Tool

Call the registered check_antibody_developability tool when chain roles are identifiable. The compute service owns implementation details and output paths; do not import a backend module or construct an output directory in the agent prompt.

Input CSV

Minimum for paired antibodies:

name,heavy_chain,light_chain,wt_name
demo_wt,HEAVY_SEQ,LIGHT_SEQ,demo_wt
demo_design,HEAVY_SEQ,LIGHT_SEQ,demo_wt

For de novo designs, wt_name can be blank when SoluProt is skipped.

Built-In Filters (when backend is available)

• pI: built in.
• Liability: built in, including CDR glycosylation and cysteine flags.
• SoluProt: skipped by the project tool.
• TAP: enabled when a compatible structure and chain mapping are provided.
• BioPhi/OASis: skipped by the project tool.

Agent Usage

When the workflow already supplies a batch result for every candidate, use those rows directly and do not repeat the compute call. Call the direct tool only when objective batch evidence is absent and a new measurement is needed.

Assess these six antibody-specific dimensions independently for every candidate:

1. Expressivity — free or unpaired Cys, long CDR hydrophobic stretches,

gross VH/VL pI mismatch, and disruption of conserved V-domain residues. Do not use codon or GC arguments when only protein sequence is supplied.

1. Immunogenicity — framework agreement with the configured scaffold,

naturalness of CDR substitutions for antibody sequence context, and unusual hydrophobic CDR peptides.

1. Aggregation — overall hydrophobic content, CDR hydrophobic patches

(especially H3), amyloid-like motifs, and unusual loop beta propensity.

1. Solubility — total and per-CDR charge, VH/VL pI mismatch, large positive

H3 patches, and extreme CDR charge.

1. Specificity — H3 positive or hydrophobic/aromatic clusters, combined-CDR

positive charge, and obvious cross-reactive motifs.

1. Liability — annotate supported hits by CDR and position: deamidation,

isomerization, N-glycosylation, Met/Trp oxidation context, unpaired Cys, and acid-cleavage DP motifs.

Use chain roles and one-based inclusive CDR ranges from the supplied antibody identification evidence. If chain identity remains uncertain, make the affected dimensions Medium Risk and explain the uncertainty.

Decision rule:

• any High Risk dimension -> overall High Risk and fail;
• otherwise any Medium Risk dimension -> overall Medium Risk and pass;
• all Low Risk dimensions -> overall Low Risk and pass.

Return one result for every supplied candidate ID and do not add candidates.

When the backend is available (available is not False)

Reconcile the tool result:

• PI_filter=fail → raise solubility/formulation concern.
• liability_filter=fail → raise liability concern.
• all_filter_pass=pass is supportive evidence, not a substitute for full reasoning.

When the backend is unavailable (result contains available: False)

The developability compute service is not installed. Apply LLM-based judgment using sequence properties:

1. pI estimation — count charged residues (Arg, Lys, His vs Asp, Glu) in

the heavy and light chains. Flag candidates likely outside pH 5–8 as formulation risks.

1. CDR liability scan — inspect CDR loops (H1, H2, H3, L1, L2, L3) for

known liability motifs: NG/NS deamidation sites, unpaired cysteines, exposed Met/Trp oxidation hotspots, and N-glycosylation sequons (N-X-S/T).

1. Sequence composition — very high hydrophobicity patches in CDR-H3 or

framework regions suggest aggregation risk.

1. Summarize findings as pi_concern, liability_concern, and

overall_developability_risk (low / medium / high) and include the reasoning in the output.

Common Mistakes

• Do not call this for target chains; it is for antibody binder chains.
• Do not omit the light chain for paired antibodies.
• Do not expect SoluProt or BioPhi/OASis unless those optional checks are enabled.
• Do not skip the assessment when available: False — fall back to LLM judgment instead.
• Do not recompute or override objective batch rows supplied by the workflow.