Skip to content

First Pass Optimize

To automatically scan for and design mutations to resolve severe developability liabilities in a single step, select one or more entries in the Project View, open the Edit menu, and select First Pass Optimize....

An interactive modal allows you to configure severe score cutoffs for AbLang, AbLang2, IgBert, and CVV (defaulting to system severe thresholds) and optionally toggle whether language model recommendations are allowed within CDR regions.

When executed, First Pass Optimize evaluates the target sequences and builds a mutation design set named "First Pass Optimize" in the Engineering workspace. If a set with that name already exists on the antibody, an auto-incremented name (e.g., First Pass Optimize 2) is assigned automatically.

Optimization Rules and Strategy

First Pass Optimize evaluates and designs substitutions using the following rules:

  • Primary Residue Focus and Flanking Evaluation: As a first pass, First Pass Optimize focuses on direct primary residue substitutions (such as \(N \rightarrow Q\) for deamidation or \(D \rightarrow E\) for isomerization). Users should evaluate flanking residues manually in the Engineering workspace when exploring alternative fix pathways or conserving key paratope contacts. Always cross-reference proposed flanking modifications with the Satlawa motif tables in Liabilities to ensure candidate flanking substitutions do not introduce secondary liabilities (e.g. mutating flanking residues in deamidation motifs to Alanine or Serine can generate medium or high deamidation motifs like \(NA\) or \(NS\)).

  • Structural PTM Gating: If a 3D structural model or PDB structure is available for the antibody, relative solvent accessibility (rASA) and sidechain SASA are calculated for all residues. Chemical PTM liabilities (deamidation, isomerization, fragmentation, hydrolysis, oxidation, and N-glycosylation) that are buried in the structural core (\(\text{rASA} < 15.0\%\) and \(\text{scASA} < 5.0\text{ \AA}^2\)) are automatically gated out from optimization since they are shielded from solvent-mediated degradation. Structural fold liabilities (such as disrupted CDR3 salt bridges, unusual or missing cysteines, and co-evolutionary or language model stability outliers) are always retained.

  • ProteinMPNN 3D Backbone Inverse Folding and Clash Flagging: When a 3D structural model or experimental PDB coordinate file is available for the antibody, First Pass Optimize evaluates candidate substitutions against ProteinMPNN conditional backbone probabilities strictly to assign strain/clash badges (Favored, Tolerated, Strained, or Clash) and log-odds scores. ProteinMPNN is not used to auto-enter or override mutation designs, preserving pure sequence-driven recommendations (language models, PFA, and chemical remediation). Candidate mutations that cause steric clash (\(\Delta\log P < -1.0\) or \(P < 0.015\)) remain active and selectable in the engineering set, but are prominently flagged with MPNN Clash warning badges and clear explanatory notes (e.g., ProteinMPNN backbone clash (ΔlogP = -1.45)...). Users who wish to adopt or explore ProteinMPNN structural recommendations can use the Suggest Alternatives... tool in the Engineering workspace.

  • Unusual Cysteine Pairing and Parental Retention Safeguards: When an Fv contains more than 2 unusual cysteines, or when an unusual cysteine is located in a predicted paratope (Paragraph contact probability \(\ge 0.50\)) or Honegger exclusion zone (when no 3D structure is available), they are mutated to Serine (S). Outside the paratope/Honegger exclusion positions (when \(\le 2\) unusual cysteines are present), unusual cysteines are evaluated as stability positions using language models, human germline PFA, and ProteinMPNN. When two or more unusual cysteines are present in the construct (forming a possible non-canonical disulfide bond), they are paired together (Together, "all or none") and all positions explicitly receive parental retention (includeParent: true / Allowed: yes). This guarantees that downstream combinatorial generation cleanly evaluates keeping the intact parental disulfide bond or remediating all cysteines concurrently, preventing asymmetric variants that leave reactive, unpaired free cysteine thiols.

  • DMS Library Auto-Pruning: If Deep Mutational Scanning (DMS) or library mutagenesis data is present on the antibody or inherited from its ancestor lineage, proposed mutations are automatically cross-referenced against the library's permissibility matrix. Unsafe mutations (\(\text{safe} = \text{false}\)) are automatically switched to alternative safe substitutions that also resolve the liability (e.g. \(N \rightarrow S/A/T\) for deamidation); if no safe replacement exists at that position, the mutation is retained as Inactive with explanatory notes.

  • Disrupted CDR3 Salt Bridge: Recommends Arginine (R) at heavy chain IMGT position 106 (if parent is not R/K) and Aspartate (D) at IMGT position 116 (if parent is not D), pairing both mutations together (Together).

  • Deamidations (N): Mutated to Glutamine (Q) or to a severe model recommendation if available at that position.

  • N-linked Glycosylation (N): Mutated to Glutamine (Q) or to a severe model recommendation (e.g. AbLang recommending S) if available at that position.

  • Isomerizations or Fragmentations (D): Mutated to Glutamic Acid (E).

  • Language Models and CVV: Selects the highest probability or frequency alternative residue based on configured cutoffs (default AbLang \(\ge 5.0\), AbLang2 \(\ge 2.0\), IgBert \(\ge 2.5\), CVV \(\ge 3.0\)). When severe model recommendations are present at a site, the top recommended residue is prioritized. If a liability model does not supply a specific replacement residue (e.g., CVV positional likelihood penalties), the residue with the highest positional frequency in the human germline repertoire (PFA) is chosen.

  • CDR Options: Excludes language model (AbLang, AbLang2, IgBert, and CVV) liabilities that fall within CDR regions by default, unless individually enabled in the modal. Region classification uses sequence-specific AntPack region annotations.

  • Target Sequence and Chain Ordering: Reads sequences directly from target antibody entries to evaluate both chains on paired Fvs. Mutated positions are presented with Light chain rows ('L') ordered before Heavy chain rows ('H').

Exclusion Areas and Parental Retention Strategy

A critical aspect of antibody optimization is managing the trade-off between eliminating liabilities and preserving antigen-binding affinity. In First Pass Optimize, the purpose of an exclusion area is specifically to set Parental Retention (includeParent: true / Allowed: yes) in the generated mutation designs:

  • Single-Choice Mutations (includeParent: false): When parental retention is not required, the design set specifies only the engineered mutant. Combinatorial variant generation forces the substitution in all built antibodies, generating a single, focused variant without expanding the design space.

  • Permissive Alternative (includeParent: true): When parental retention is required, both the engineered mutant and the parental wild-type residue are permitted as options during combination generation. This ensures you can evaluate variants carrying the mutation alongside variants that preserve the wild-type residue if the site is sensitive for binding.

Structure-Based Paratope Predictions vs. Static Honegger Exclusions

Historically, First Pass Optimize relied on a static sequence mask derived from Honegger humanization exclusions combined with Vernier zone positions. While effective for sequence-only workflows, a static sequence mask designates over half of all typical candidate positions as requiring wild-type retention—even for framework residues, Vernier positions facing away from antigen, and solvent-exposed loops with zero contact propensity. This causes unnecessary exponential growth in variant combinations (\(2^N\)).

To provide precision engineering, First Pass Optimize dynamically switches its exclusion strategy based on structural availability:

  • When a 3D Structure Exists (Paratope-then-Honegger Hierarchy): If a 3D structural model (e.g., from ABodyBuilder2) or experimental structure is available, FPO queries residue-level paratope predictions generated by the Paragraph Equivariant Graph Neural Network (EGNN).

    • Calibrated Contact Threshold (\(\ge 0.50\)): Paragraph outputs an antigen-binding probability score between \(0.00\) and \(1.00\). The \(0.50\) threshold is the validated decision boundary separating non-contact positions from bona fide antigen-binding contacts.

    • True Paratope and Loop Residues (\(\text{Score} \ge 0.50\) or CDR Loops): Any position with contact propensity \(\ge 0.50\) or located in a CDR loop is identified as an active binding interface residue and receives includeParent: true (Allowed: yes). This guarantees that the wild-type residue remains available during combinatorial screening.

    • Non-Contact Framework Residues (\(\text{Score} < 0.50\)): Scaffold residues with contact propensity below \(0.50\) in framework regions do not require wild-type retention and receive includeParent: false (Allowed: -). This eliminates combinatorial inflation while confidently repairing developability liabilities.

    • VHH Interface Mask: For single-chain VHH antibodies, hallmark framework interface positions (IMGT 42, 49, 50, and 52) continue to require parental retention to protect solubility and autonomous folding.

  • When No 3D Structure Exists (Honegger Fallback): If the antibody entry lacks 3D coordinate data, First Pass Optimize automatically falls back to sequence-based exclusions. Positions falling within CDR regions, Vernier zones, Honegger humanization exclusion areas, or VHH hallmark positions are assigned includeParent: true.

Downstream Combinatorial Generation

Once First Pass Optimize builds the design set:

  1. Open the Engineering workspace for the target antibody.
  2. Select the "First Pass Optimize" set from the Set dropdown.
  3. Review the mutation rows, noting the (P) badge in the Region cell for residues confirmed as predicted paratopes.
  4. Filter by Min Mutations and Max Mutations if desired, or use Generate Combinations to build discrete variants for downstream developability profiling.