After studying this material, you should be able to:

1. Explain why conventional antibody discovery is slow and costly
2. Describe the three critical factors in antibody drug development
3. Understand how AI models address each factor: binding affinity ranking, developability prediction, and de novo design
4. Identify key technical innovations in MochiBind, CA-MAP, and the agent-guided design pipeline
5. Recognize how experimental validation integrates with computational design in a lab-in-the-loop workflow

---

Step-by-Step Teaching

Step 1: The Problem — Why Antibody Discovery Needs AI

What Are Monoclonal Antibodies?

Monoclonal antibodies (mAbs) are laboratory-made proteins that mimic the immune system's ability to target specific threats, such as cancer cells or pathogens. They are among the most successful drug types, with over 100 FDA-approved drugs.

The Core Problem

Despite their success, traditional discovery is:

• ⏱️ Slow: 6–12 months from target to lead candidate
• 💰 Expensive: Requires extensive lab work and iteration
• 🔁 Inflexible: Hard to adapt when targets change

The Three Questions Every Antibody Drug Must Answer

🔑 Key Insight: Existing AI models perform well on familiar targets but poorly on new ones. Benchmarks have hidden this gap, making it hard to measure and fix.

---

Step 2: Factor 1 — Ranking Binders (MochiBind)

The Challenge

You have thousands of antibody candidates. Most don't bind at all. You need to find the ones worth testing in the lab — without running expensive experiments on all of them.

Why Old Methods Fall Short

Previous models tried to predict absolute binding affinity (an exact number). This fails because:

• Numbers aren't comparable across different labs or assays
• Models trained on known targets perform poorly on new ones
• Test sets rarely include non-binders, making evaluations misleadingly optimistic

MochiBind's Solution: Pairwise Ranking

Instead of predicting an exact number, MochiBind asks:

"Between antibody A and antibody B targeting the same antigen — which binds more tightly?"

How it works, step by step:

Step 1: Represent antibody-antigen complexes
        ↓
        Use ESM-2 (protein language model) to embed residues
        ↓
Step 2: Compress each complex into one embedding
        ↓
        Average all residue embeddings → single vector
        ↓
Step 3: Project into lower-dimensional space
        ↓
        Specialized network layer compares two candidates
        ↓
Step 4: Aggregate pairwise comparisons
        ↓
        TrueSkill algorithm (from video game ranking) → global ranking

🎮 Analogy: TrueSkill was originally used to rank Xbox players based on match wins/losses. Here it ranks antibodies based on pairwise binding comparisons.

Why This Is Better

Evaluation Framework

The team used the AlphaBind dataset covering four antigens (TIGIT, PD-1, HER2, SARS-CoV-1 RBD) with ~30,000 variants each. The strict rule: train on some antigens, test on a completely different one — simulating real-world use on novel targets.

---

Step 3: Factor 2 — Predicting Developability (CA-MAP)

What Is Developability?

A drug candidate must not only bind its target — it must also:

• ✅ Stay stable in the bloodstream
• ✅ Not clump together (aggregate)
• ✅ Not trigger an immune response
• ✅ Survive manufacturing processes

These properties are collectively called developability.

The Core Problem: Batch Effects

Batch effect: Systematic measurement differences between labs due to different equipment, protocols, or handling — not biology.

When a model trains on one lab's data, it silently learns that lab's biases. When applied to a different lab's data, predictions become unreliable.

CA-MAP's Solution: Context-Aware Prediction

CA-MAP (Context-Aware Multi-property Antibody Predictor) uses a prompt-based approach:

Input Prompt Structure:
┌─────────────────────────────────────────┐
│ Example antibody 1 + measured property  │ ← From same lab as query
│ Example antibody 2 + measured property  │ ← Captures batch offset
│ Example antibody 3 + measured property  │
│ ...                                     │
│ Query antibody + property name?         │ ← What we want to predict
└─────────────────────────────────────────┘

The examples from the same lab carry the batch offset information. The model learns to read and correct for it automatically.

The Training Trick: AB-Context-Aware

The problem: a model might just ignore the examples and rely only on the query sequence.

Solution: During training, apply a random hidden transformation to both the context measurements and the expected answer. The model cannot recover the correct answer without reading the context — so it's forced to use it.

Performance Comparison

ρ = Spearman correlation; 1.0 = perfect prediction

Architecture Efficiency

Bonus Capability: Predicting Unseen Properties

CA-MAP can estimate properties it was never directly trained on by leveraging correlations between developability properties:

• Trained on 4 properties → tested on 2 new ones
• Immunogenicity: ρ improved from 0.25 → 0.73 when correlated properties were included in context
• This means expensive assays could be estimated from cheaper ones

---

Step 4: Factor 3 — Designing Antibodies From Scratch

The Campaign: A Real Cancer Target

The team targeted a protein on desmoplastic small round-cell tumors — a rare, aggressive pediatric cancer. This target was chosen because:

• It sits on tumor cell surfaces (accessible)
• Driven by a specific genetic error
• Largely absent from healthy tissue (selective)
• No existing structure, no prior antibody data — a true test of de novo design

Phase 1: Where to Attack? (Hotspot Recommendation Agent)

Before designing antibodies, you must decide which region of the target to aim for (the epitope/hotspot).

The agent orchestrates 7 bioinformatics tools analyzing:

Target Protein
     ↓
┌─────────────────────────────────────┐
│ Tool 1: Solvent-accessible surface  │
│ Tool 2: Secondary structure         │
│ Tool 3: Hydrophobicity              │
│ Tool 4: Sequence uniqueness         │
│ Tool 5: Immune Epitope Database     │ ← 500,000 entries
│ Tool 6: Protein family annotation   │
│ Tool 7: Negative target avoidance   │
└─────────────────────────────────────┘
     ↓
Agent synthesizes → Ranked hotspot recommendations
                    with biophysical rationale

Performance: Recovered at least one true epitope residue in top 5 recommendations ~80% of the time on diverse holdout targets.

Phase 2: Generate Antibody Designs

Three generative models with different approaches each produced 96,000 designs:

Phase 3: Score and Filter Candidates

Each design was scored on:

• 🔬 Folding confidence: Will it fold into the right shape?
• 🔗 Complex quality: Will it form the correct binding interface?
• ⚠️ Sequence liabilities: Any manufacturing/development red flags?
• 💪 Binding affinity: MochiBind's ranking score

A candidate selection agent applied multi-objective Pareto filtering — keeping designs that excel on different combinations of metrics, not just one.

📊 Pareto filtering: Retains candidates where no other candidate is better on all metrics simultaneously — preserving diversity of strong solutions.

100,000 candidates were prioritized for experimental screening.

Phase 4: Experimental Validation

100,000 computational designs
         ↓
    Yeast display screening
    (synthesized, displayed on yeast surface)
         ↓
    Two rounds of sorting/filtering
         ↓
    116 survivors — none bound unrelated control protein
    (confirming target specificity, not general stickiness)
         ↓
    Individual binding measurement
         ↓
    46 confirmed strong binders ✅

Phase 5: Lab-in-the-Loop

The 46 binders + all binder/non-binder labels become training data for the next design cycle:

Design → Score → Experiment → Learn → Better Design → ...

Each round of experiments sharpens the models for the next round.

---

Step 5: Putting It All Together

The Complete AI-Driven Antibody Discovery Pipeline

TARGET PROTEIN
      ↓
[Hotspot Agent] → Identifies where to bind
      ↓
[Generative Models] → Creates 288,000 candidate designs
      ↓
[MochiBind] → Ranks by predicted binding strength
      ↓
[CA-MAP] → Predicts developability properties
      ↓
[Pareto Filtering] → Selects diverse top candidates
      ↓
[Lab Validation] → Experimental confirmation
      ↓
[Lab-in-the-Loop] → Data feeds back to improve models

Summary of Key Innovations

---

Key Takeaways

1. Antibody drug discovery has three core challenges: finding the right binding site, ranking binders, and predicting developability
2. Generalization to novel targets is the hardest problem — models must be evaluated cross-antigen, not just cross-variant
3. Relative prediction (ranking) is more robust than absolute prediction across different experimental contexts
4. Batch effects can destroy model performance; context-aware architectures can correct for them at inference time
5. Computational design + experimental validation in a feedback loop is more powerful than either alone
6. Speed matters: 200,000 pairs in 13 seconds enables screening at scales impossible with wet-lab methods alone