AI Antibody Design: From Target Antigen to Therapeutic Candidate in One Afternoon

The $300 Billion Antibody Therapeutics Market

Monoclonal antibodies are the fastest-growing class of therapeutics in pharmaceutical history. In 2025, antibody therapeutics generated over $300 billion in global revenue, led by drugs like adalimumab (Humira), pembrolizumab (Keytruda), and nivolumab (Opdivo). Over 100 antibodies are approved by the FDA, with hundreds more in clinical trials across oncology, autoimmune disease, infectious disease, and neurology. The specificity, potency, and modularity of antibodies make them ideal drug scaffolds – they can be engineered to bind virtually any extracellular target with picomolar affinity.

Despite their clinical success, antibody discovery remains slow and expensive. The traditional workflow starts with immunizing mice with the target antigen, generating hybridoma cell lines, screening thousands of clones for binding, and then humanizing the best hits to reduce immunogenicity in patients. This process takes 6-12 months and costs $200,000-500,000 per campaign. Phage display libraries offer an alternative but still require weeks of panning, screening, and affinity maturation. Both approaches are fundamentally empirical – they generate random diversity and screen for the desired property.

AI is changing this equation. Structure prediction tools like ImmuneBuilder can predict antibody 3D structures with sub-angstrom accuracy on CDR loops. Generative design tools like AntiFold can propose new CDR sequences optimized for target binding, stability, and humanness. The combination enables a workflow that goes from target antigen to therapeutic candidates in a single afternoon – compressing months of empirical screening into hours of computation.

This guide walks through the complete AI antibody design pipeline: from target antigen structure to candidate antibodies ready for experimental validation. We use a real-world example – designing anti-PD-L1 antibodies – and provide working Python code for every step using the SciRouter SDK.

Traditional Antibody Discovery: Why It Takes So Long

To appreciate what AI changes, it helps to understand the traditional workflow. Hybridoma technology, developed by Kohler and Milstein in 1975 (Nobel Prize, 1984), involves immunizing a mouse with the target protein, harvesting splenic B cells, and fusing them with myeloma cells to create immortalized antibody-producing clones. Each clone produces a unique monoclonal antibody. The researcher then screens thousands of clones by ELISA or flow cytometry to find those that bind the target with high affinity and specificity.

The bottleneck is screening. A typical hybridoma campaign produces 10,000-50,000 clones, of which only 0.1-1% will have the desired binding properties. Screening takes 2-4 weeks of labor-intensive assays. Once binders are identified, their variable regions must be sequenced, reformatted as recombinant antibodies, re-expressed, and characterized for affinity, specificity, stability, and developability. The entire process from immunization to characterized lead typically takes 6-9 months.

Phage display (Nobel Prize, 2018, George Smith) is faster because it bypasses animal immunization. Synthetic or natural antibody variable region libraries are displayed on phage particles and panned against immobilized antigen. After 3-4 rounds of selection (2-3 weeks), enriched clones are sequenced and characterized. However, phage display libraries may not contain high-affinity binders for every target, and the selected clones often require affinity maturation – additional rounds of mutagenesis and selection to improve binding from micromolar to nanomolar or picomolar affinity.

AI design addresses the fundamental limitation of both approaches: they search sequence space randomly. An antibody has approximately 120 variable region residues, with the six CDR loops comprising about 60 positions. Even considering only the 20 natural amino acids, CDR sequence space is astronomical (20^60 possible sequences). Random screening explores a vanishingly small fraction. Computational design narrows the search to sequences that are structurally compatible with the target, dramatically increasing the fraction of functional binders.

The AI Antibody Design Pipeline

The SciRouter antibody design pipeline chains four steps: (1) obtain or predict the antigen structure, (2) predict the antibody template structure with ImmuneBuilder, (3) design new CDR sequences with AntiFold, and (4) validate designs by refolding and checking humanness. Each step takes seconds to minutes via API, and the entire pipeline runs in under an hour even for complex targets.

Step 1: Prepare the Antigen Structure

You need the 3D structure of your target antigen to guide antibody design. If an experimental structure exists (check the PDB at rcsb.org), use it directly. For novel targets, predict the structure withESMFold or Boltz-2. The antigen structure defines the epitope – the surface region where the antibody will bind – and informs CDR design.

For our case study, we use PD-L1 (Programmed Death-Ligand 1), a key immune checkpoint target in cancer immunotherapy. PD-L1 binds PD-1 on T cells, suppressing anti-tumor immunity. Blocking this interaction with antibodies (atezolizumab, durvalumab, avelumab) restores T cell function. The crystal structure of PD-L1 is available as PDB 4ZQK, but we will predict it from sequence to demonstrate the full workflow.

from scirouter import SciRouter

client = SciRouter(api_key="sk-sci-YOUR_KEY")

# PD-L1 extracellular domain (human, UniProt Q9NZQ7, residues 19-239)
pdl1_sequence = (
    "FTVTVPKDLYVVEYGSNMTIECKFPVEKQLDLAALIVYWEMEDKNIIQFVHGEEDLKVQ"
    "HSSYRQRARLLKDQLSLGNAALQITDVKLQDAGVYRCMISYGGADYKRITVKVNAPYNKI"
    "NQRILVVDPVTSEHELTCQAEGYPKAEVIWTSSDHQVLSGKTTTTNSKREEKLFNVTSTL"
    "RINTTTNEIFYCTFRRLDPEENHTAELVIPELPLAHPPNERTHLVILGA"
)

pdl1_fold = client.proteins.fold(sequence=pdl1_sequence)
print(f"PD-L1 folded: pLDDT = {pdl1_fold.mean_plddt:.1f}")
pdl1_pdb = pdl1_fold.pdb_string

Step 2: Predict Antibody Template Structure with ImmuneBuilder

ImmuneBuilder is a deep learning model specifically trained to predict antibody 3D structures. Unlike general-purpose protein structure predictors (AlphaFold2, ESMFold), ImmuneBuilder is optimized for the unique features of antibody variable regions – particularly the hypervariable CDR loops that are critical for antigen binding. It achieves sub-angstrom backbone RMSD on CDR-H3, the most variable and functionally important loop.

You provide the heavy and light chain variable region sequences, and ImmuneBuilder returns a predicted Fv structure. This structure serves as the template for CDR redesign in the next step. We start with a well-characterized anti-PD-L1 antibody framework: atezolizumab (Tecentriq), whose variable region sequences are publicly available from patent filings and crystal structures (PDB: 5XXY).

# Atezolizumab-like variable regions as starting template
heavy_chain = (
    "EVQLVESGGGLVQPGGSLRLSCAASGFTFSDSWIHWVRQAPGKGLEWVAWISPYGGSTYY"
    "ADSVKGRFTISADTSKNTAYLQMNSLRAEDTAVYYCARRHWPGGFDYWGQGTLVTVSS"
)

light_chain = (
    "DIQMTQSPSSLSASVGDRVTITCRASQDVSTAVAWYQQKPGKAPKLLIYSASFLYSGVPS"
    "RFSGSGSGTDFTLTISSLQPEDFATYYCQQYLYHPATFGQGTKVEIKR"
)

# Predict the Fv structure
ab_structure = client.antibodies.fold(
    heavy_chain=heavy_chain,
    light_chain=light_chain,
)
print(f"Antibody folded: {len(ab_structure.pdb_string)} bytes")
print(f"CDR-H3 confidence: {ab_structure.cdr_h3_confidence:.2f}")

ImmuneBuilder identifies the six CDR loops automatically using IMGT numbering, reports per-residue confidence scores, and highlights the CDR-H3 loop which is typically the primary determinant of binding specificity. High CDR-H3 confidence (above 0.8) indicates the loop conformation is well-predicted and provides a reliable template for redesign.

Step 3: Design New CDR Sequences with AntiFold

AntiFold is an inverse folding model specifically trained on antibody structures. Given an antibody backbone structure (from ImmuneBuilder), AntiFold proposes new CDR sequences that are predicted to maintain the overall fold while potentially improving binding affinity, stability, or humanness. It is the antibody-specific equivalent of ProteinMPNN, optimized for the unique structural features of immunoglobulin variable domains.

AntiFold can redesign individual CDR loops or all six CDRs simultaneously. For most therapeutic applications, you redesign CDR-H3 (the primary specificity determinant) while keeping the other CDRs and framework regions fixed. This focused approach maintains the overall antibody architecture while exploring diversity in the most functionally important region.

The temperature parameter controls sequence diversity, similar to ProteinMPNN. Low temperature (0.1) produces conservative variants close to the input sequence. Higher temperature (0.2-0.4) generates more diverse CDR-H3 sequences that explore broader binding modes. For initial discovery, a temperature of 0.2-0.3 provides a good balance between diversity and fold quality.

# Design 50 CDR-H3 variants
design_result = client.antibodies.design(
    pdb_string=ab_structure.pdb_string,
    design_cdrs=["H3"],       # Focus on CDR-H3
    num_sequences=50,
    temperature=0.25,
    fix_framework=True,        # Preserve framework regions
)

print(f"Designed {len(design_result.sequences)} CDR-H3 variants")
for i, seq in enumerate(design_result.sequences[:10]):
    print(f"  Variant {i+1}:")
    print(f"    CDR-H3: {seq.cdr_h3_sequence}")
    print(f"    Score: {seq.score:.3f}")
    print(f"    Length: {len(seq.cdr_h3_sequence)} residues")

Step 4: Validate and Rank Candidates

Designed CDR sequences must be validated computationally before experimental testing. The validation pipeline checks three properties: (1) fold quality – does the designed antibody fold correctly when re-predicted by ImmuneBuilder? (2) humanness – how similar is the framework to the closest human germline? (3) developability – are there aggregation-prone motifs, deamidation sites, or unpaired cysteines that would cause manufacturing problems?

# Validate all designs by refolding with ImmuneBuilder
validated = []

for seq in design_result.sequences:
    # Reconstruct full heavy chain with new CDR-H3
    new_heavy = seq.full_heavy_chain

    # Refold to check structural integrity
    refold = client.antibodies.fold(
        heavy_chain=new_heavy,
        light_chain=light_chain,
    )

    # Check CDR-H3 confidence and overall fold quality
    if refold.cdr_h3_confidence > 0.7:
        validated.append({
            "cdr_h3": seq.cdr_h3_sequence,
            "heavy_chain": new_heavy,
            "design_score": seq.score,
            "h3_confidence": refold.cdr_h3_confidence,
            "pdb": refold.pdb_string,
        })

validated.sort(key=lambda x: x["design_score"])
print(f"{len(validated)} designs pass fold validation")
print(f"\nTop 5 candidates:")
for i, v in enumerate(validated[:5]):
    print(f"  {i+1}. CDR-H3: {v['cdr_h3']}")
    print(f"     Score: {v['design_score']:.3f}, "
          f"H3 confidence: {v['h3_confidence']:.2f}")

Designs that pass fold validation (CDR-H3 confidence above 0.7) are then checked for humanness by comparing framework sequences to the IMGT database of human germline V, D, and J genes. Frameworks with greater than 85% identity to the closest human germline are considered acceptably human for therapeutic development. AntiFold preserves framework residues by default, so humanness issues are rare when starting from a humanized template.

Case Study: Designing Anti-PD-L1 Antibody Candidates

PD-L1 (Programmed Death-Ligand 1) is one of the most validated targets in immuno-oncology. Three anti-PD-L1 antibodies are FDA-approved: atezolizumab (Roche/Genentech), durvalumab (AstraZeneca), and avelumab (Pfizer/Merck KGaA). Each blocks the PD-1/PD-L1 interaction from the ligand side, preventing immune checkpoint engagement and restoring T cell anti-tumor activity.

Despite three approved drugs, there is ongoing interest in next-generation anti-PD-L1 antibodies with improved affinity, pH-dependent binding (for enhanced FcRn recycling), bispecific formats, and reduced immunogenicity. AI design can rapidly explore these variations. Let's run the full pipeline, starting from the atezolizumab framework and designing diverse CDR-H3 variants.

# Full pipeline: anti-PD-L1 antibody design

# 1. Fold PD-L1 antigen (already done above)
print(f"Antigen ready: PD-L1 pLDDT = {pdl1_fold.mean_plddt:.1f}")

# 2. Fold template antibody
template = client.antibodies.fold(
    heavy_chain=heavy_chain,
    light_chain=light_chain,
)
print(f"Template ready: CDR-H3 confidence = {template.cdr_h3_confidence:.2f}")

# 3. Design 100 CDR-H3 variants at two temperatures
designs_conservative = client.antibodies.design(
    pdb_string=template.pdb_string,
    design_cdrs=["H3"],
    num_sequences=50,
    temperature=0.15,
    fix_framework=True,
)

designs_exploratory = client.antibodies.design(
    pdb_string=template.pdb_string,
    design_cdrs=["H3"],
    num_sequences=50,
    temperature=0.35,
    fix_framework=True,
)

all_designs = designs_conservative.sequences + designs_exploratory.sequences
print(f"Generated {len(all_designs)} total CDR-H3 variants")

# 4. Validate all by refolding
candidates = []
for seq in all_designs:
    refold = client.antibodies.fold(
        heavy_chain=seq.full_heavy_chain,
        light_chain=light_chain,
    )
    if refold.cdr_h3_confidence > 0.75:
        candidates.append({
            "cdr_h3": seq.cdr_h3_sequence,
            "score": seq.score,
            "confidence": refold.cdr_h3_confidence,
        })

candidates.sort(key=lambda x: x["score"])
print(f"\n{len(candidates)} candidates pass validation")
print(f"Ready for experimental expression and binding assays")

In a real campaign, the validated candidates would be ordered as synthetic gene fragments from a vendor like Twist Bioscience or IDT (approximately $100-200 per gene for antibody variable regions). The genes are cloned into expression vectors, expressed in HEK293 or CHO cells, purified by Protein A chromatography, and tested for binding to PD-L1 by surface plasmon resonance (SPR) or bio-layer interferometry (BLI). The entire experimental validation phase takes 2-4 weeks and costs $5,000-15,000 for 20 candidates.

Compare this to the traditional hybridoma approach: 3-4 months of mouse immunization, 2 months of hybridoma fusion and screening, 2-3 months of humanization and reformatting. Total cost: $200,000-500,000. Total timeline: 8-12 months. The AI pipeline produces experimentally validated candidates in 3-5 weeks at a fraction of the cost.

Understanding CDR Design: What AntiFold Optimizes

AntiFold operates on the inverse folding principle: given a desired 3D backbone conformation, predict amino acid sequences that will adopt that conformation. For antibodies, this means designing CDR sequences that maintain the loop geometry required for antigen recognition while optimizing for overall fold stability.

The model considers several structural features when proposing CDR sequences. Residues at the base of CDR loops must form proper hydrogen bonds with the framework to anchor the loop. Residues in the loop apex (the part that contacts the antigen) can be more diverse because they are solvent-exposed and less constrained by the fold. Internal loop residues must maintain the loop's canonical structure through intra-loop hydrogen bonds and hydrophobic contacts.

CDR-H3 is special because it lacks canonical structures. While CDR-H1, H2, L1, L2, and L3 adopt a limited set of known conformations (canonical classes), CDR-H3 is structurally diverse and hard to predict. This is both a challenge and an opportunity: CDR-H3 encodes the most binding specificity information, and its structural diversity means there are many possible solutions for any given target. AntiFold handles this diversity by conditioning its predictions on the local structural environment rather than relying on canonical class templates.

When designing all six CDRs simultaneously (rather than just CDR-H3), AntiFold maintains the VH-VL interface – the critical contact zone between heavy and light chains that holds the Fv fragment together. Disrupting this interface leads to poor expression and aggregation. The model has learned these interface constraints from the thousands of antibody structures in its training set.

Humanization Assessment

For any antibody intended for therapeutic use in humans, humanness is a critical quality attribute. Non-human framework sequences can trigger anti-drug antibodies (ADAs) that neutralize the therapeutic and cause adverse reactions. The FDA requires immunogenicity assessment for all therapeutic antibodies, and high humanness scores reduce the risk of ADA formation.

Humanness is quantified by comparing the antibody variable region sequence to human germline V, D, and J gene segments from the IMGT database. The closest human germline hit and the percent identity to that germline provide a humanness score. Antibodies with greater than 85% framework identity to the closest human germline are generally considered acceptably humanized. Fully human antibodies (from transgenic mice or human phage libraries) typically show greater than 95% germline identity.

Because AntiFold designs CDR sequences while preserving framework regions, starting from a humanized or fully human template (like atezolizumab) ensures that the designed antibodies maintain high humanness. The CDR sequences themselves do not have germline references (they are somatically hypermutated in nature), so CDR humanness is assessed differently – primarily by checking for unusual amino acid compositions, deamidation-prone Asn-Gly motifs, and oxidation-sensitive unpaired methionines.

# Check humanness of top candidates
for candidate in candidates[:5]:
    heavy = candidate.get("heavy_chain", heavy_chain)

    # Framework identity to closest human germline
    # (simplified - real implementation uses IMGT BLAST)
    human_residues = sum(
        a == b for a, b in zip(heavy[:25], "EVQLVESGGGLVQPGGSLRLSCAAS")
    )
    framework_identity = human_residues / 25

    # Check for problematic motifs in CDR-H3
    h3 = candidate["cdr_h3"]
    has_ng = "NG" in h3     # Deamidation risk
    has_dp = "DP" in h3     # Isomerization risk
    has_met = "M" in h3     # Oxidation risk

    print(f"CDR-H3: {h3}")
    print(f"  Framework humanness: {framework_identity:.0%}")
    print(f"  Deamidation risk (NG): {'Yes' if has_ng else 'No'}")
    print(f"  Isomerization risk (DP): {'Yes' if has_dp else 'No'}")
    print(f"  Oxidation risk (Met): {'Yes' if has_met else 'No'}")
    print()

Using the Antibody Design Studio

The Antibody Design Studio provides a visual interface for the entire pipeline. Enter heavy and light chain sequences, and the studio automatically predicts the antibody structure with ImmuneBuilder, displays it with CDR loops highlighted, and sets up the design workflow.

Interactive CDR Selection

The 3D structure viewer highlights all six CDR loops in distinct colors. Click on individual CDRs to toggle them for redesign. By default, only CDR-H3 is selected. The studio shows the current CDR sequences alongside their confidence scores, helping you identify loops that might benefit from redesign (lower confidence often indicates suboptimal sequences).

Design Parameters

Configure the number of variants (10-100), sampling temperature (0.1-0.5), and whether to fix framework regions. The studio provides presets for common scenarios: "Conservative optimization" (temperature 0.15, single CDR), "Diverse exploration" (temperature 0.35, single CDR), and "Full redesign" (temperature 0.25, all CDRs). Each preset includes recommended parameters based on published benchmarks.

Results and Export

Designed variants appear in a sortable table with CDR sequences, design scores, fold confidence after refolding, and developability flags (deamidation, isomerization, oxidation motifs). Click any variant to overlay its predicted structure with the template in the 3D viewer. Export results as FASTA for gene synthesis or CSV for downstream analysis. The studio also generates a summary report suitable for inclusion in patent applications and regulatory filings.

Beyond Single Antibodies: Bispecifics and Nanobodies

The AI antibody design pipeline extends beyond conventional monoclonal antibodies. Bispecific antibodies – engineered to bind two different targets simultaneously – are a rapidly growing therapeutic modality with over 15 approved drugs as of 2026. Designing bispecifics requires optimizing two sets of CDRs simultaneously, ensuring that both binding arms fold correctly and do not interfere with each other. AntiFold can design CDRs for each arm independently, and ImmuneBuilder can predict the structure of each Fv domain to check for steric clashes.

Nanobodies (VHH domains from camelid heavy-chain-only antibodies) are another important format. Nanobodies are smaller (15 kDa vs 150 kDa for IgG), more stable, easier to produce in microbial systems, and can access epitopes inaccessible to conventional antibodies. ImmuneBuilder supports nanobody structure prediction through its dedicated NanoBodyBuilder2 module. AntiFold can design CDR3 loops for nanobodies using the same inverse folding approach.

The SciRouter Antibody Design Studio supports all three formats: conventional IgG (heavy + light chain), bispecific Fv domains, and nanobodies (single VHH domain). Select the format before starting the design workflow, and the studio adjusts the pipeline accordingly.

From Computation to Clinic: What Happens Next

Computational design produces candidate sequences. Experimental validation turns candidates into drugs. The typical path from AI-designed antibody to clinical candidate involves several stages, each with decreasing numbers of candidates and increasing investment.

• Gene synthesis and expression (20 candidates): Order synthetic genes, clone into expression vectors, express in HEK293 or CHO cells. Timeline: 2–3 weeks. Cost: $5,000–$15,000.
• Binding characterization (20 candidates): Measure affinity to target by SPR (Biacore) or BLI (Octet). Filter for sub-nanomolar KD. Timeline: 1–2 weeks. Cost: $3,000–$8,000.
• Functional assays (5–10 candidates): Test for blocking activity, cell-based potency, species cross-reactivity. Timeline: 2–4 weeks. Cost: $10,000–$30,000.
• Developability assessment (3–5 candidates): Thermal stability (DSF), aggregation (SEC-MALS), viscosity, polyreactivity. Timeline: 2–3 weeks. Cost: $10,000–$20,000.
• Lead selection (1–2 candidates): Choose the best overall candidate for IND-enabling studies. Total elapsed time from AI design to lead: 2–3 months.

Compare this 2-3 month timeline (at $30,000-75,000 total cost) to the traditional 8-12 month timeline (at $200,000-500,000). AI antibody design does not eliminate experimental validation, but it dramatically reduces the time and cost by starting from computationally optimized candidates rather than random library screening. The hit rate for AI-designed antibodies in binding assays is typically 30-50%, compared to 1-5% for naive library selections, meaning fewer candidates need to be tested to find a lead.

The convergence of AI structure prediction, generative CDR design, and automated experimental platforms (robotic antibody expression, high-throughput SPR) is creating a new paradigm for therapeutic antibody discovery. What once required a team of 10 scientists working for a year can now be accomplished by a computational biologist with an API key and a contract research organization – in weeks, not months, at a fraction of the traditional cost.