The Language of Life (Part 4): 'Dreaming Up' Antibodies—Generative AI for Biologics

Table of Contents

• The Biologics Challenge: Beyond Small Molecules
• Humanome.ai's In Silico Biologics Design
  • Stage 1: Designing the "Business End" (CDRs)
• The Real Challenge: "Developability"
• Co-Generation of Binding and Developability
• Conclusion

The Biologics Challenge: Beyond Small Molecules

We have covered de novo proteins (Part 2) and de novo small molecules (Part 3). Now we turn to biologics—specifically, therapeutic antibodies.

Antibodies are the fastest-growing and one of the most effective classes of drugs, valued for their exquisite specificity. But as any biologics team knows, they are "messy". They are massive, complex proteins.

The "Old Way" of discovering them is slow, expensive, and analog:

• Immunization: Discovering them in mice and then "humanizing" them
• Library Screening: Screening vast phage display libraries

Both methods are resource-intensive. Worse, the "humanization" process—grafting the mouse-derived binding loops (CDRs) onto a human scaffold—is a "patch." It is a slow, error-prone process that often fails, resulting in candidates with high immunogenicity (the tendency to be rejected by the human immune system).

Humanome.ai's In Silico Biologics Design

We skip this entire messy process. We design de novo antibodies in two stages, entirely in silico.

Stage 1: Designing the "Business End" (CDRs)

An antibody's binding specificity is defined by its six Complementarity-Determining Region (CDR) loops. The target's binding surface is the "epitope."

Using the 3D structure of our target epitope (the "lock"), we use our generative models (from Part 2, such as RFdiffusion or SOTA binder-design models like OriginFlow) to de novo design CDRs that are a perfect complementary match.

This single in-silico step allows us to generate a library of 10^6 or more antibody variants, all specifically "dreamed up" to bind the target epitope, often with predicted affinities higher than those found in nature.

The Real Challenge: "Developability"

This is the central problem in biologics. A perfect binder is not a drug. The primary reason biologics fail in development is "poor developability".

We define "Developability" as a set of critical, non-negotiable biophysical properties:

• Low Immunogenicity: Does it look "foreign" to the human immune system?
• High Solubility: Will it aggregate (clump together) in a high-concentration vial?
• Thermostability: Does it denature and unfold at physiological temperatures?
• Manufacturability: Can it be expressed and purified at scale?

Co-Generation of Binding and Developability

The "Old Way" is painfully sequential:

1. Find a binder (12 months)
2. Test its developability (6 months)
3. Watch it fail (e.g., it's insoluble)
4. Go back to Step 1

This multi-year, billion-dollar cycle is the bottleneck. Our platform shatters this with multi-objective co-generation.

Just as we did with small molecules in Part 3, we optimize for all properties simultaneously at the moment of creation. The "Old Way" is sequential filtering; our way is simultaneous, guided generation.

SOTA research explicitly describes a "guidance approach" that integrates this property information directly into the diffusion model's sampling process. This allows the simultaneous design of CDRs (for binding) "while considering critical properties like solubility and folding stability".

Our generative process is a multi-modal, guided process. As our diffusion model generates the antibody sequence and structure, it is guided at each step by a panel of predictive "developability" models.

Our multimodal "Developability Prompt" looks like this:

Generate(Antibody) WHERE reward = 
  (w1 * Binding_Affinity_Score) + 
  (w2 * SITA_Immunogenicity_Score) + 
  (w3 * SAbPred_TAP_Score) + 
  (w4 * CamSol_Solubility_Score)

We are not just designing binders; we are designing drug candidates. We optimize for low immunogenicity, high stability, and solubility from the first amino acid. This means the candidates that come out of our in-silico platform are already de-risked and optimized for manufacturing and clinical success.

Conclusion

We have now designed a novel protein (Part 2), a small molecule (Part 3), and an antibody. But these are all, for now, just hypotheses. How do we test them at scale?

Next in Part 5, we will show how we test all 1,000 designs in one day in our "Virtual Lab."

#antibodyDesign #biologics #developability #therapeuticAntibodies #CDRdesign