Antibody development — end-to-end in silico map

Self-contained HTML: diagrams are static scalable vector graphics (SVG) (no internet or JavaScript required).

Phase 0: Problem framing

Diagram (nodes and edges)

Scientist thinking (hypotheses and setup)

High-level hypothesis: Good outcomes depend on correct target context, well-defined objectives, and consistent data schemas; downstream results are only as reliable as the framing and inputs.
Target context: Record isoform and construct choice, oligomeric state and binding partners, post-translational modifications (PTM) (for example glycosylation), and native environment assumptions.
Schema conventions: Define residue numbering and complementarity determining region (CDR) boundary conventions so every tool uses compatible indices.
Bookkeeping: Decide what will be stored per candidate (scores, flags, complex models, paratope masks, and decisions) before running large-scale screening.

Candidate “biography” data structure (example)

candidate c:
    epitope_identifier: [E_i]
    origin_program: [Program 2A (assembly), Program 2B (generative)]
    variable_heavy_chain_sequence, variable_light_chain_sequence
    three_dimensional_complex_model
    paratope_protection_mask
    scores:
        binding: {per_tool_scores, ensemble_score}
        developability: {therapeutic_antibody_profiler, aggrescan3d, camsol, stability_proxy}
        safety: {cross_reactivity_docking, motif_mimicry}
    risk_flags: [String]
    probability_of_success: P(success | evidence)

Phase 1: Epitope discovery

Diagram (nodes and edges)

Scientist thinking (hypotheses and experiments)

Epitope-level hypothesis: Epitopes that are exposed in the native context, conserved, and functionally relevant are more likely to yield effective binders and less likely to fail later.
Cheap first, expensive later: Run permissive static prediction and contextual filters on a broad pool, then run molecular dynamics only on a shortlist.
Experiment 1A: Permissive static epitope prediction using Experiment 1A-1 (BepiPred), Experiment 1A-2 (Ellipro), and Experiment 1A-3 (Discotope) to generate a broad candidate pool.
Experiment 1B: Contextual epitope filters: accessibility and masking, conservation and escape risk, functional site overlay, and optional off-target similarity screening.
Experiment 1C: Conditional molecular dynamics simulation for top candidates or when static predictors disagree.
Experiment 1D: Evidence integration and an explicit epitope gate; failures loop back to adjust filters or expand the pool.

Tools, inputs, outputs, and parameters

Tool or step	Input	Output	Parameters and thresholds
BepiPred Epitope identification (sequence-based)	Protein sequence	Per-residue epitope probability score and predicted linear epitope segments	Epitope threshold: Higher threshold increases precision (typically fewer residues predicted)
Ellipro Epitope identification (structure-based clustering)	Protein three-dimensional (3D) structure	Clusters of residues forming linear or conformational epitopes	Minimum score: Higher values are more stringent and return fewer epitopes Maximum distance (angstrom (Å)): Clustering radius for discontinuous epitopes
Discotope Epitope identification (structure-based scoring)	Protein three-dimensional (3D) structure	Per-residue epitope probability score	Structure confidence filter: Optionally ignore low-confidence regions
Accessibility and masking analysis (for example solvent accessible surface area (SASA) and glycan masking rules) Contextual epitope filter	Epitope patches on one or more conformational states; optional glycan model and complex partners	Accessibility score per epitope; masking flags (glycan shielding, steric occlusion); exposure frequency across states	Accessibility threshold: Minimum solvent exposure required Masking rules: What counts as occluded in the native context State weighting: How to combine exposure over conformations
Conservation and escape-risk screen (multiple sequence alignment) Contextual epitope filter	Target sequence set (homologs, variants, or strains) and epitope residue indices	Conservation score per epitope; predicted escape-risk annotation	Sequence set definition: Which homologs or variants to include Conservation cutoff: Minimum acceptable conservation Escape weighting: How strongly to penalize variable residues
Functional site annotation overlay Contextual epitope filter	Known or predicted functional residues (active sites, receptor-binding interfaces, allosteric sites)	Functional relevance score per epitope and rationale notes	Functional distance threshold: How close an epitope must be to a functional region Evidence weighting: How to treat curated versus predicted annotations
Off-target surface-patch similarity screen (optional) Early cross-reactivity risk screen	Epitope patch descriptors and an off-target surface library	Similarity hits and a pre-design cross-reactivity risk flag	Similarity metric: Shape and physicochemical similarity definition Hit threshold: Similarity cutoff for flagging Library scope: Which off-targets to include
Molecular dynamics simulation (for example GROMACS, AMBER) Dynamic interrogation (expensive; shortlist only)	Static structure and a shortlist of epitopes or regions	Trajectory; residue flexibility metrics (root mean square fluctuation (RMSF)); dominant conformational states	Simulation length: Approximately 100–500 nanoseconds (or longer if needed) Solvent model: Explicit solvent box Selection rule: Run only on top N epitopes or when predictors disagree
Evidence integration and epitope gate Decision step	All epitope scores and annotations (static, contextual, dynamic, functional)	Ranked epitope list; explicit pass or fail decision with reason codes	Weighting scheme: Combine static consensus, accessibility, conservation, functional relevance, and dynamics Hard constraints: Any non-negotiable criteria (for example must be exposed)

Phase 2: Paratope design

Diagram (nodes and edges)

Scientist thinking (hypotheses and programs)

Paratope-level hypothesis: For each selected epitope, designs with plausible antibody–antigen complexes and diverse binding modes are more likely to succeed than a large set of near-duplicates.
Cheap first, expensive later: Generate many candidates, cluster and pre-filter cheaply, then do complex modeling and constrained refinement on a shortlist.
Complex evidence is explicit: Binding scores come from explicit complex modeling and interface plausibility checks (not from refinement alone).
Paratope protection mask: Store interface residues so Phase 3 optimizations (for example solubility design) can avoid breaking binding.
Gate and loops: If binding is implausible, diversity collapses, or immunogenicity risk is high, loop back to redesign rather than proceeding forward.

Tools, inputs, outputs, and parameters

Tool or step	Input	Output	Parameters and thresholds
OptMAVEn-2.0 Paratope design (knowledge-based module assembly)	Antigen three-dimensional (3D) structure with epitope residues and a database of variable heavy chain (VH) and variable light chain (VL) modules	Ranked variable heavy chain and variable light chain designs (sequences and assembled structures) with binding energies or scores	Epitope definition: Patch definition used for design Module libraries: Germline and framework constraints Energy thresholds: Cutoffs for retaining designs
RosettaAntibodyDesign (RAbD) Paratope design (complementarity determining region redesign)	Antigen structure with target epitope region and an antibody framework for complementarity determining region grafting	Ranked designs as Protein Data Bank format (PDB) structures with new complementarity determining region sequences and structures	Complementarity determining regions to design: Choose among light chain complementarity determining regions 1–3 (L1–L3) and heavy chain complementarity determining regions 1–3 (H1–H3) Interface options: Penalize unsatisfied polar atoms; enforce hydrogen bonds and salt bridges
De novo generative design (diffusion model or graph neural network (GNN)) Paratope design (generative)	Epitope three-dimensional coordinates and a fixed antibody framework	Large sets of de novo complementarity determining region loop structures and sequences	Number of generated designs: Typically thousands per epitope Initial filters: Plausibility, diversity, and internal energy
Diversity management and clustering Down-selection step (cheap first)	Design set (sequences and structures) from one or more programs	Cluster representatives covering distinct sequences and binding modes; redundancy-reduced set	Clustering metric: Sequence identity and structural similarity Diversity targets: Minimum number of clusters or modes to retain
Cheap pre-filters (geometry and sequence sanity checks) Down-selection step (cheap first)	Candidate antibody models and frameworks	Filtered set removing obvious steric clashes and out-of-distribution sequences	Clash thresholds: Maximum allowed internal clashes Framework constraints: Allowed frameworks and complementarity determining region length bounds
Complex modeling and interface scoring Binding evidence production	Candidate antibodies, epitope structures, and optional conformational ensembles	Antibody–antigen complex poses; binding scores; interface plausibility metrics	Pose count: Number of complex poses retained per candidate Interface checks: Clash cutoff; buried unsatisfied polar groups penalty Scoring function: Which binding score or ensemble score is used
AbDesign Constrained refinement (expensive; shortlist only)	Antigen structure and antibody backbone fragments or ensembles	Refined antibody models with optimized interface and stability metrics	Docking constraints: Key epitope residues and desired interactions Iteration budget: Number of refinement rounds
Humanization and immunogenicity screen (optional) Risk-reduction gate	Candidate sequences and germline reference sets; major histocompatibility complex (MHC) class II peptide risk models	Humanization recommendations; immunogenicity risk flags	Germline proximity target: Maximum divergence allowed Peptide-risk thresholds: Cutoffs and allele set used
Paratope definition artifact (protection mask) Interface bookkeeping	Complex model(s) for each candidate	List of interface residues to protect during Phase 3 optimization (for example during CamSol design mode)	Interface definition: Distance cutoff for defining contact residues Ensemble rule: Union or intersection across poses
Paratope gate Decision step	Binding score evidence, diversity statistics, and immunogenicity flags	Pass or fail decision with reason codes and loops back to design if needed	Binding plausibility thresholds: Minimum interface quality requirements Diversity requirements: Minimum cluster coverage Immunogenicity constraints: Hard fail versus flag rules

Phase 3: Developability and safety

Diagram (nodes and edges)

Scientist thinking (hypotheses and gates)

Developability hypothesis: Strong predicted binders are viable only if they pass sequence liabilities, aggregation risk, solubility, stability proxies, and safety screens.
Parallel, not serial: Most screens are independent and can run in parallel, feeding a single gate with explicit aggregation rules.
Additional liabilities: Include chemical liabilities, variable-region glycosylation motifs, and self-interaction, polyspecificity, and viscosity risk proxies alongside aggregation and solubility.
Fix liabilities and re-score loop: Failures trigger mutation or variant proposals (protecting the paratope) and re-scoring; if the interface changes, return to Phase 2 refinement.
Decision clarity: The gate should distinguish “flag” from “fail” and state how docking and motif screens combine into a final safety decision.

Tools, inputs, outputs, and parameters

Tool or step	Input	Output	Parameters and thresholds
Therapeutic Antibody Profiler (TAP) Developability (sequence liabilities)	Variable heavy chain (VH) and variable light chain (VL) sequences	Pass, flag, or fail status against core sequence metrics with residue drivers	Structure usage: Whether to model or use a provided structure Flag versus fail policy: How many flags are tolerated before failing
Chemical liability scan Developability (sequence liabilities)	Candidate sequences	Liability sites (for example deamidation, isomerization, oxidation) and suggested mitigations	Motif definitions: Which sequence patterns to flag Context filters: Structural exposure requirements, if used
Glycosylation motif risk scan Developability (sequence liabilities)	Candidate sequences	Potential glycosylation motifs in variable regions and risk flags	Motif definition: How N-linked motifs are detected Region scope: Which regions are evaluated (for example complementarity determining regions only)
Self-interaction, polyspecificity, and viscosity risk proxy Developability (formulation and in vivo risk proxies)	Candidate sequences and or structures	Predicted self-interaction or nonspecific binding risk, plus viscosity risk proxies, with flags and drivers	Proxy model choice: Sequence-only versus structure-informed predictor Thresholds: Risk cutoffs used for gating
Aggrescan3D Developability (aggregation risk in three-dimensional context)	Candidate structure in Protein Data Bank format (PDB); optional chain or region selection	Aggregation-prone surface patches and suggested point mutations	Exposure and sphere radius: Around carbon alpha (Cα) for exposure calculations Mutation scan options: Which residues and substitutions to try
CamSol Developability (solubility prediction and optimization)	Candidate sequences and an optional paratope protection mask from Phase 2	Solubility scores and ranked mutation suggestions or variant libraries	Regions to protect: Protect the paratope during design Mutation budget: Number of mutations allowed per variant Environment assumptions: pH and ionic strength
Stability proxy (for example thermal stability prediction) Developability (stability)	Candidate sequences and or structures	Stability score proxy and risk flags	Model choice: Which stability proxy is used Cutoffs: Minimum acceptable stability proxy value
AntiTarget-Dock Safety (cross-reactivity docking screen)	Candidate antibody models and a structural library of high-risk proteins	Predicted off-target binding energies and a docking-based risk score	Anti-target library: Which proteins are included Docking thresholds: Binding energy or score cutoff for flagging or failing
Proto-Blast-Search (Basic Local Alignment Search Tool for proteins (BLASTp)) Safety (motif mimicry screen)	Candidate sequences and a reference proteome database	Motif similarity matches and a motif-based risk score contribution	Motif length and thresholds: Similarity cutoff and permitted mismatches Aggregation rule: How motif and docking risk combine into pass, flag, or fail
Risk aggregation and developability gate Decision step	Outputs of all developability and safety screens	Pass or fail decision; prioritized liability list; triggers a fix-and-re-score loop on failure	Hard-fail thresholds: Non-negotiable cutoffs per screen Weighted aggregation: How multiple risks are combined Safety aggregation: How docking and motif screens are thresholded and combined

Phase 4: Integrated analysis and learning

Diagram (nodes and edges)

Scientist thinking (ranking and learning loop)

Separate prediction from selection: The probability of success model produces calibrated probabilities; multi-objective ranking then applies developability and safety constraints and trade-offs.
Multi-objective ranking: Use a Pareto-style view so candidates are compared across probability, developability, and safety rather than collapsed into a single score without explanation.
Explorer candidates: Select a small number of high-uncertainty candidates for expected information gain, not just “one or two by habit.”
Measurement model: Replicates, assay variance, controls, and batch handling determine how much the outcomes should update calibration.
Feedback: Outcomes update priors and tool weights, and can also trigger changes in Phase 0 objectives and constraints when reality disagrees with assumptions.

Tools and decision steps

Tool or step	Input	Output	Parameters and thresholds
Candidate dossier synthesis Evidence packaging	All candidate artifacts (epitope rationale, complex models, scores, flags)	Per-candidate dossier suitable for review and decision making	Dossier fields: What is included and how it is versioned
Probability of success model Calibration and prediction	In silico features and prior wet lab outcomes (if any)	Calibrated probability estimates per candidate	Calibration method: Platt scaling, isotonic regression, or Bayesian calibration Feature set: Which in silico signals are included
Multi-objective ranking (Pareto selection) Decision support	Probability estimates plus developability and safety evidence	Shortlist based on trade-offs (Pareto-efficient set) and constraints	Constraint definition: What must be satisfied before ranking Trade-off policy: How to handle competing objectives
Wet lab selection policy with expected information gain Experiment design	Ranked shortlist plus uncertainty estimates	Final candidate set to test: exploit for success and explore for learning	Selection budget: Number of candidates that can be tested Explorer rule: How expected information gain or uncertainty is quantified
Measurement model and data quality plan Reliability plan	Assay definitions and operational constraints	Replicate plan; controls; batch randomization; data normalization strategy	Replicate count: Number of repeats per candidate Batch policy: Randomization and correction approach
Feedback and updating Learning loop	Wet lab outcomes and metadata	Updated calibration, tool weights, and design priors; potential updates to Phase 0 constraints	Update schedule: When and how models are re-fit Change management: How updated priors and constraints are recorded

Antibody development — end-to-end map (tools, outputs, parameters)

Overall pipeline

Glossary of common terms

Phase 0: Problem framing

Phase 1: Epitope discovery

Phase 2: Paratope design

Phase 3: Developability and safety

Phase 4: Integrated analysis and learning