AI in High-Throughput Antibody Discovery: Integrating De Novo Design with Surface Plasmon Resonance

Paradigm Shift in Therapeutic Biologics: Epitope Diversity vs Affinity #

In therapeutic antibody discovery, developmental campaigns historically over-indexed on nanomolar or picomolar equilibrium dissociation constants (K_D). High-affinity binders were prioritized early in screening cascades, often through repeated rounds of selection. However, clinical drug development demonstrates that epitope targeting—the exact spatial and topological contact residues on the antigen surface—is the primary determinant of therapeutic mechanism of action (MoA).

                      THE THERAPEUTIC ANTIBODY TRIAD
                 ┌───────────────────────────────────────┐
                 │                EPITOPE                │
                 │   Functional Geometry & Mechanism:    │
                 │ Agonism, Antagonism, Allostery, ADCC  │
                 └──────────────────┬────────────────────┘
                                    │
           ┌────────────────────────┴────────────────────────┐
           ▼                                                 ▼
┌──────────────────────┐                          ┌──────────────────────┐
│       AFFINITY       │                          │    DEVELOPABILITY    │
│  Binding Kinetics    │                          │ Solubility, Viscosity│
│  (kon, koff, KD)     │                          │ Aggregation, PK/PD   │
└──────────────────────┘                          └──────────────────────┘

An antibody with moderate affinity directed against a functional neutralizing epitope can be affinity-matured downstream via targeted mutagenesis. Conversely, an antibody with picomolar affinity targeting a non-neutralizing or immunologically silent epitope cannot be engineered into an effective therapeutic without altering its binding site.

Furthermore, comprehensive epitope binning provides critical intellectual property (IP) protection. Patent offices increasingly reject claims based purely on sequence complementarity-determining regions (CDRs) or target binding, requiring functional boundary definitions established via competitive epitope blocking matrices.

---

Biophysical Mechanisms of Epitope Binning #

Epitope binning is an empirical competitive cross-blocking assay wherein monoclonal antibodies (mAbs) are tested pairwise against an antigen to determine whether they compete for the same, overlapping, or distinct steric footprints.

                           STERIC VS ALLOSTERIC BLOCKING
       Direct Steric Overlap                        Allosteric Conformational Shift
  ┌─────────────────────────────┐              ┌─────────────────────────────┐
  │         [Antigen]           │              │         [Antigen]           │
  │        /         \          │              │        /         \          │
  │     [mAb 1]     [mAb 2]     │              │    [mAb 1]    Conformational │
  │   Shared spatial binding    │              │    Binding     Restructuring│
  │   residues collide.         │              │    Induces ──> Prevents mAb2 │
  └─────────────────────────────┘              └─────────────────────────────┘

Competitive behavior arises through two distinct biophysical mechanisms:

1. Steric Hindrance: The two antibodies share identical contact residues or bind adjacent epitopes whose physical molecular volumes (each IgG molecule having a molecular weight of ~150 kDa and dimensions of 14.5 nm x 8.5 nm x 4.0 nm) mutually exclude simultaneous occupancy.
2. Allosteric Perturbation: Binding of the first antibody induces a conformational shift in the antigen tertiary structure, destabilizing or eliminating the distant binding pocket required by the second antibody.

The Three Fundamental Biosensor Formats #

                           THE THREE BINNING ARCHITECTURES
     1. In-Tandem (Classical)          2. Premix Format           3. Array-Based Co-Injection
  ┌─────────────────────────────┐ ┌─────────────────────────┐ ┌─────────────────────────────────┐
  │ Sensor ──> [mAb 1 Capt.]    │ │ Pre-incubate Antigen    │ │ Sensor ──> [Target Capture]     │
  │ Step 1 ──> Inject Antigen   │ │ with mAb 1 in solution; │ │ Microfluidic 96-printhead       │
  │ Step 2 ──> Inject mAb 2     │ │ Inject over immobilized │ │ captures 384 distinct mAbs in   │
  │ Detect: Additive vs Blocked │ │ mAb 2 biosensor.        │ │ continuous array matrix.        │
  └─────────────────────────────┘ └─────────────────────────┘ └─────────────────────────────────┘

1. In-Tandem (Classical Sandwich): mAb 1 is covalently coupled or captured to the sensor surface. Antigen is injected to saturation. Subsequently, mAb 2 is injected. An increase in mass/refractive index demonstrates simultaneous binding (non-competing; distinct epitopes). An absence of binding signifies cross-blocking (same bin).
2. Premix Format: Antigen is pre-incubated in solution with a molar excess of mAb 1 until equilibrium is established. The mixture is then flowed over a sensor chip presenting immobilized mAb 2. If the complex fails to bind the surface, mAb 1 and mAb 2 compete for the target epitope.
3. Array-Based Multiplexed Co-Injection: Enabled by high-throughput continuous flow microfluidics (such as Carterra LSA), this approach prints an array of hundreds of antibodies onto a single planar chip, flowing antigens and sandwiching antibodies across the array in an automated sequence.

---

Instrumental Architectures: HT-SPR vs Bio-Layer Interferometry (BLI) #

The evolution from low-throughput, dual-channel biosensors to high-throughput platforms has enabled large-scale competitive matrices in early discovery:

---

Machine Learning & AI-Driven Epitope Discovery #

High-throughput biophysical binning generates structured matrices that feed machine learning models, moving antibody engineering from empirical screening toward predictive design.

                   AI-DRIVEN ANTIBODY DISCOVERY CYCLE
  [HT-SPR 384x384 Matrix] ──> [Network Graph Clustering] ──> [AlphaFold3 Structural Modeling]
   Unbiased Competition       MCL Community Detection        Epitope Paratope Interface
                                                                         │
                                                                         ▼
  [Wet-Lab Functional Assay] <── [In Silico De Novo Design] <── [RFdiffusion / ESM-2]
   Validated Hit Candidate        Targeted CDR Generation         Protein Language Models

1. Large Protein Language Models (pLMs) #

Protein language models (ESM-2, AntiBERTa, AbLang) pre-trained on billions of evolutionary protein sequences encode internal representations of structural constraints and biochemical fitness:

• Zero-Shot Binding Likelihood: Predicts whether mutating a specific CDR-H3 residue preserves or disrupts binding to a target epitope without requiring crystal structures.
• Humanization & Developability: Models evaluate human-likeness scores (BioPhi, Sapiens) to optimize developability metrics (preventing hydrophobic patch aggregation and non-specific polyspecificity) while maintaining epitope fidelity.

2. De Novo Structural Generation (AlphaFold3 & RFdiffusion) #

• RFdiffusion: Employs diffusion probabilistic generative models to scaffold functional CDR loops de novo. By specifying target epitope coordinates on an antigen crystal or cryo-EM structure, RFdiffusion designs de novo backbone geometries that contact the target site.
• AlphaFold3-Multimer: Generates complex predictions between polyclonal antibody repertoires and multi-domain antigens, predicting interface root-mean-square deviation (iRMSD) to identify potential shared-binding epitopes prior to synthesis.

3. Graph Community Detection Algorithms (In Silico Binning) #

Raw competitive blocking matrices are parsed using graph theory algorithms:

• Markov Cluster Algorithm (MCL): Simulates random walks across a weighted graph where nodes represent individual mAbs and edges represent competitive blocking probabilities. Flow simulation clusters antibodies into discrete functional communities (bins).
• Leiden & Louvain Network Clustering: Optimizes modularity scores to identify hierarchical sub-bins (e.g., distinguishing antibodies that share an overlapping core footprint from those with non-identical peripheral interactions).

---

Step-by-Step Bench SOP: 384-Array Epitope Binning on Carterra LSA #

                           CARTERRA LSA BENCH PROTOCOL
  [Chip Activation] ──> [Array Printing] ──> [Ethanolamine Quench] ──> [Antigen Prime]
   EDC / NHS Solution    384 mAbs via 96-Tip   1.0M Block Unreacted    10-50 nM Analyte
                                                                               │
                                                                               ▼
  [Regeneration Wash] <── [Heatmap Analysis] <── [Second mAb Injection] <──────┘
   10 mM Glycine pH 2.0    Graph Network Binning   Sandwich Competition

Reagent Specifications & Sensor Preparation #

• Sensor Chip: Carterra HC30M (High-Capacity Polycarboxylate Hydrogel, 30 nm hydrogel thickness).
• Coupling Chemistry: 100 mM 1-ethyl-3-(3-dimethylaminopropyl)carbodiimide (EDC) mixed 1:1 with 100 mM N-hydroxysuccinimide (NHS) in 10 mM MES (pH 5.5).
• Blocking Buffer: 1.0 M Ethanolamine-HCl (pH 8.5).
• Running / Assay Buffer: HBS-EP+ (10 mM HEPES, 150 mM NaCl, 3 mM EDTA, 0.05% v/v Surfactant P20, pH 7.4), supplemented with 0.1 mg/mL bovine serum albumin (BSA) to eliminate non-specific surface adsorption.

Step 1: Microfluidic Immobilization of 384 Antibody Print Array #

1. Equilibrate the HC30M chip in running buffer for 30 minutes. Dock into the Carterra LSA instrument at 20°C.
2. Prime the 96-channel microfluidic printhead with deionized water, followed by running buffer.
3. Activate the chip surface across the printhead array by injecting freshly prepared EDC/NHS solution for 7 minutes at 10 µL/min.
4. Print 384 individual monoclonal antibodies (diluted to 5–10 µg/mL in 10 mM Sodium Acetate buffer, pH 4.5–5.0) in four consecutive 96-channel printing cycles. Contact time: 10 minutes per block to achieve capture densities of 800 to 1,500 Resonance Units (RU).
5. Deactivate unreacted NHS esters across the entire chip surface by injecting 1.0 M Ethanolamine-HCl (pH 8.5) across the single flow cell for 7 minutes.

Step 2: Automated Classical Sandwich Screening Cascade #

Operate the instrument in single-flow-cell (SFC) mode to circulate solutions across all 384 array spots simultaneously:

Cycle Architecture for Each Analyte mAb (Repeated across all 384 antibodies):
  1. Baseline Step:       Flow Running Buffer for 60 seconds (Establish R_baseline).
  2. Antigen Capture:     Inject Target Antigen (20 nM in Running Buffer) for 240 seconds.
  3. Intermediate Buffer: Flow Running Buffer for 60 seconds (Verify stable antigen baseline).
  4. Sandwich mAb Influx: Inject Candidate mAb (10 µg/mL) for 240 seconds (Measure R_sandwich).
  5. Dissociation Step:   Flow Running Buffer for 90 seconds.
  6. Surface Regeneration:Inject Regeneration Cocktail (10 mM Glycine-HCl, pH 2.0) for 45 seconds.
  7. Stabilization:       Flow Running Buffer for 120 seconds.

Repeat this sequential injection series for all 384 antibodies in the panel to populate the complete pairwise interaction matrix.

---

Data Analysis, Heatmap Thresholding & Network Graph Bins #

                      REPRESENTATIVE 8x8 EPITOPE BLOCKING HEATMAP
        mAb-1   mAb-2   mAb-3   mAb-4   mAb-5   mAb-6   mAb-7   mAb-8
  mAb-1 [  B  ] [  B  ] [  B  ] [  -  ] [  -  ] [  -  ] [  -  ] [  -  ] ──> BIN 1 (Red)
  mAb-2 [  B  ] [  B  ] [  B  ] [  -  ] [  -  ] [  -  ] [  -  ] [  -  ] ──> BIN 1 (Red)
  mAb-3 [  B  ] [  B  ] [  B  ] [  -  ] [  -  ] [  -  ] [  -  ] [  -  ] ──> BIN 1 (Red)
  mAb-4 [  -  ] [  -  ] [  -  ] [  B  ] [  B  ] [  -  ] [  -  ] [  -  ] ──> BIN 2 (Blue)
  mAb-5 [  -  ] [  -  ] [  -  ] [  B  ] [  B  ] [  -  ] [  -  ] [  -  ] ──> BIN 2 (Blue)
  mAb-6 [  -  ] [  -  ] [  -  ] [  -  ] [  -  ] [  B  ] [  B  ] [  B  ] ──> BIN 3 (Green)
  mAb-7 [  -  ] [  -  ] [  -  ] [  -  ] [  -  ] [  B  ] [  B  ] [  B  ] ──> BIN 3 (Green)
  mAb-8 [  -  ] [  -  ] [  -  ] [  -  ] [  -  ] [  B  ] [  B  ] [  B  ] ──> BIN 3 (Green)
        Legend: [ B ] = Cross-Blocked (Shared Epitope); [ - ] = Sandwich Formed (Distinct Epitope)

Signal Processing & Normalization #

1. Reference Subtraction: Subtract signals from blank interstitial reference spots to correct for bulk refractive index shifts, microfluidic pressure changes, and non-specific binding to the dextran matrix.
2. Normalized Binding Calculation: Calculate the Binding Metric (BM) to categorize pairwise interactions:

Categorization Thresholds: #

• BM < 0.15 (Blocked): Complete steric or allosteric cross-blocking. Both mAbs belong to the same epitope bin.
• BM > 0.40 (Not Blocked): Simultaneous sandwich binding confirmed. mAbs target distinct, non-overlapping epitopes.
• 0.15 ≤ BM ≤ 0.40 (Intermediate / Partial Blocking): Indicates partial steric clash, lower binding affinity, or allosteric negative cooperativity.

Bidirectional Symmetry Validation #

Every interaction pair [i, j] is tested in both configurations:

• Configuration A: mAb i captured o Antigen o mAb j sandwich.
• Configuration B: mAb j captured o Antigen o mAb i sandwich.

Discrepancies where Pair [i, j] is blocked but Pair [j, i] sandwiches (asymmetry) often reflect differences in binding affinities, avidity effects on homodimeric antigens, or allosteric dissociation kinetics.

---

Technical Troubleshooting Matrix for HT Epitope Binning #

---

Normative Guidelines & Literature Citations #

1. Abdiche, Y. N., et al. (2014). High-throughput epitope binning of large antibody libraries using surface plasmon resonance imaging (SPRi). MAbs, 6(1), 195–208.
2. Brooks, B. D., et al. (2021). High-throughput epitope binning using the Carterra LSA platform: A comprehensive review of protocols and data analysis. Methods, 180, 48–60.
3. Abanades, B., et al. (2023). ImmuneBuilder: Deep-Learning models for predicting the 3D structures of immune proteins. Communications Biology, 6(1), 575.
4. Watson, J. L., et al. (2023). De novo design of protein structure and function with RFdiffusion. Nature, 620(7976), 1089–1100.
5. International Organization for Standardization (ISO). (2020). ISO 20387:2018 Biotechnology — Biobanking — General requirements for biobanking.