Deep learning assessment of nativeness and pairing likelihood for antibody and nanobody design with AbNatiV2

This paper introduces AbNatiV2 and its paired variant p-AbNatiV2, deep learning models trained on millions of sequences to significantly improve the assessment of nativeness and pairing likelihood for both nanobodies and conventional antibodies, thereby supporting more effective design and humanization strategies.

The Gist

The Big Picture: The "Immune System's DNA"

Imagine the human immune system as a massive, highly skilled orchestra. For millions of years, it has been composing perfect songs (antibodies) that fight diseases without causing chaos (toxicity) or playing the wrong notes (attacking the body's own cells).

Scientists want to write their own "songs" (design new medicines) that sound just like the immune system's hits. But when they try to compose these songs from scratch using computers, they often end up with "noise"—sequences that look okay on paper but fall apart in the real world or cause side effects.

To fix this, the authors of this paper created AbNatiV2, a new "musical ear" (AI tool) that can listen to a protein sequence and instantly tell: "Does this sound like a natural immune system song, or does it sound fake?"

They also built a new feature called p-AbNatiV2, which doesn't just listen to solo instruments; it checks if the Heavy and Light chains (the two main parts of an antibody) actually like each other and fit together perfectly.

---

Part 1: The Problem with the Old Tool (AbNatiV1)

The authors previously built a tool called AbNatiV1. Think of it as a music critic who was very good at judging human songs. However, it had two big flaws:

1. It was tone-deaf to "Nano-songs": It struggled to judge "nanobodies" (tiny, single-piece antibodies found in camels and llamas) because it hadn't heard enough of them.
2. It only listened to solos: It judged the Heavy and Light chains separately. In reality, an antibody is a duet. If you pick a great singer for the Heavy part and a great singer for the Light part, but they hate each other, the song will fail. The old tool couldn't see that.

Part 2: The Upgrade (AbNatiV2)

The team went back to the drawing board and built AbNatiV2. Here is how they upgraded the "critic":

• The Massive Library: They went to the library of life and found 20 million new nanobody sequences (from camels, llamas, and alpacas) that they had never seen before. They fed this massive library into the AI.
  • Analogy: If AbNatiV1 studied 2 million books, AbNatiV2 studied 20 million. It now knows the "dialect" of camel antibodies perfectly.
• The Better Brain: They upgraded the AI's architecture. Instead of just memorizing patterns, it now understands context.
  • Analogy: Imagine a chef who used to just taste ingredients. Now, the chef understands how the salt, the heat, and the timing interact to create a perfect dish. The new AI understands how amino acids interact with each other across the whole protein.
• The "Focus" Filter: They taught the AI to pay extra attention to the "spicy" parts of the protein (the variable regions) rather than the boring, repetitive parts. This helps it spot subtle differences that make a protein natural or artificial.

The Result: AbNatiV2 is now a master judge. It can tell the difference between a real camel nanobody and a computer-generated fake with incredible accuracy. It can also tell you if you've taken a "CDR" (the part of the antibody that grabs the virus) and pasted it onto a new body, and if that graft looks suspicious or natural.

---

Part 3: The New Feature (p-AbNatiV2) – The "Couples Counselor"

This is the biggest breakthrough. The team realized that for standard human antibodies, the Heavy and Light chains must be a perfect match.

They built p-AbNatiV2, which acts like a dating app for antibodies.

• How it works: You give it a Heavy chain and a Light chain.
• The Question: "Do these two belong together?"
• The Training: They trained it on 3.7 million real, naturally paired human antibody sequences.
• The Test: They tried to trick it by mixing a Heavy chain from one person with a Light chain from a completely different person (or even a different species).
  • The Result: The old tools (Humatch, ImmunoMatch) got confused and said, "Eh, they might work." But p-AbNatiV2 said, "Nope, that's a mismatch."
  • In tests, p-AbNatiV2 correctly identified the "native couple" (the real pair) 74% of the time, beating all previous models.

Why Does This Matter?

Imagine you are building a bridge.

• Old Way: You build the left side of the bridge and the right side separately, making sure they look good. Then you try to bolt them together. Sometimes they don't fit, and the bridge collapses.
• New Way (AbNatiV2): You use a simulator that checks if the left and right sides are designed to fit together before you even build them.

Real-world applications:

1. Better Drugs: Scientists can design new antibodies that are less likely to cause allergic reactions in humans (because they sound more "human").
2. Saving Time: Instead of making 1,000 antibodies in a lab and testing them all (which takes months), they can use this AI to filter out the 900 that will likely fail.
3. Nano-Medicine: It makes it much easier to engineer those tiny camel antibodies for complex tasks, like targeting cancer cells or crossing the blood-brain barrier.

The Bottom Line

The authors have released this tool (AbNatiV2 and p-AbNatiV2) as free software and a website. They are essentially giving the scientific community a super-powered compass that points toward "natural, safe, and stable" antibody designs, helping to speed up the creation of life-saving medicines.

Technical Summary

1. Problem Statement

Antibody and nanobody (VHH) therapeutics require a delicate balance between high binding affinity, stability, and low immunogenicity (nativeness/humanness). While generative models can create vast libraries of sequences, many lack "nativeness"—the likelihood of belonging to natural immune repertoires—leading to poor biophysical properties or in vivo failure.

• Limitations of Previous Work (AbNatiV1): The original AbNatiV model was effective for nanobodies but constrained by a small, low-diversity training dataset. Crucially, it operated on unpaired sequences, making it unsuitable for conventional antibodies (VH-VL pairs) where heavy and light chain interactions are critical for structural integrity and function.
• The Gap: There was a need for a unified framework that could:
  1. Assess nanobody nativeness with higher accuracy using larger, more diverse datasets.
  2. Evaluate the "humanness" of conventional antibody heavy and light chains.
  3. Predict the pairing likelihood of VH-VL pairs to ensure stable Fv regions and avoid mis-pairing in synthetic libraries or multispecific formats.

2. Methodology

The authors introduced AbNatiV2 (for unpaired sequences) and p-AbNatiV2 (for paired sequences), both based on a Vector-Quantized Variational Auto-Encoder (VQ-VAE) architecture with Transformer components.

A. Dataset Curation

• Nanobodies (VHH): Expanded the training corpus from 2 million to 21.2 million sequences. This included new deep sequencing of non-immune alpaca libraries (9.2M sequences) and curated data from recent literature (INDI dataset, etc.).
• Human Antibodies: Curated ~19.6M VH and ~21.2M VL unpaired sequences from the Observed Antibody Space (OAS).
• Paired Human Antibodies: Constructed a dataset of 3.7 million unique paired VH-VL sequences by combining p-IgGen and PairedAbNGS datasets.

B. Model Architecture & Innovations

• AbNatiV2 (Unpaired):
  • Architecture: Up-scaled Transformer-based VQ-VAE (92M parameters vs. 17M in v1).
  • Key Components: Rotary positional embeddings, gated attention mechanisms, and SwiGLU activation functions.
  • Loss Function: Introduced a Focal Reconstruction Loss (FRL). Unlike standard Mean Squared Error (MSE), FRL dynamically down-weights residues the model reconstructs with high confidence (conserved framework regions). This forces the model to focus learning capacity on variable regions (CDRs) and somatic mutations, capturing high-order dependencies better.
  • Training: Self-supervised masked learning.
• p-AbNatiV2 (Paired):
  • Architecture: A Cross-Attention VQ-VAE. It takes pre-trained unpaired VH and VL models and interconnects them via cross-attention layers in both the encoder and decoder.
  • Training Strategy: The weights of the pre-trained unpaired models are frozen; only the cross-attention layers and a new Pairing Prediction Head are trained.
  • Pairing Head: Uses Noise-Contrastive Learning (NCL). It classifies native pairs (positives) against two types of negatives:
    1. Shuffled pairs: Light chains randomly reassigned within the same batch (hard negatives).
    2. Mismatched pairs: Chains from different species, B-cell classes, or studies (easy negatives).
  • Output: Provides a joint humanness score, individual chain scores, residue-level profiles, and a specific pairing likelihood score.

3. Key Results

A. Nanobody Nativeness (AbNatiV2)

• Performance: Achieved a PR-AUC of 0.984 in distinguishing native camelid sequences from artificial PSSM-generated sequences, compared to 0.961 for AbNatiV1.
• Generalization: Significantly improved performance on "diverse" test sets (sequences >5% different from training data), where AbNatiV1 failed (PR-AUC dropped from 0.926 to 0.977).
• CDR Grafting: Successfully detected nativeness loss when CDRs were grafted onto a universal scaffold. AbNatiV2 identified 90% of grafting-induced nativeness drops, compared to 64% for AbNatiV1.

B. Human Antibody Humanness

• Unpaired: The unified VL model (training on both $\kappa$ and $\lambda$) outperformed separate v1 models, particularly in distinguishing human from rhesus sequences (PR-AUC 0.949 vs. 0.808/0.851).
• Paired: p-AbNatiV2 achieved near-perfect classification (PR-AUC $\approx$ 1.0) for human paired sequences against non-human (mouse/rat) sequences.

C. Pairing Likelihood Prediction (p-AbNatiV2)

• Accuracy: In 1-vs-1 comparisons (native pair vs. random shuffled pair with similar germline identity), p-AbNatiV2 assigned a higher score to the native pair in 74.3% of cases.
• Ranking: In 1-vs-50 scenarios, the native pair was ranked in the top 5 in >35% of cases.
• Comparison: Substantially outperformed state-of-the-art competitors Humatch (60% accuracy) and ImmunoMatch (60% accuracy).
• Robustness: Unlike ImmunoMatch, which assigned high scores to artificial PSSM sequences and mouse pairs (out-of-distribution failure), p-AbNatiV2 correctly assigned low scores to these non-human/artificial inputs.

D. Therapeutic Application

• The models successfully mapped the "nativeness landscape" of clinical-stage nanobodies, showing that successful humanized nanobodies (e.g., Caplacizumab) occupy a specific trade-off space between VHH-nativeness and VH-humanness.
• Correlation with clinical Anti-Drug Antibody (ADA) immunogenicity data was comparable to the best existing tools (Pearson $r \approx -0.5$).

4. Significance and Impact

• Unified Framework: AbNatiV2 and p-AbNatiV2 provide the first comprehensive toolset to assess nativeness and pairing for both single-domain nanobodies and conventional paired antibodies.
• Engineering Pipeline: The models enable:
  • Hit Selection: Filtering synthetic libraries for developable candidates.
  • Humanization: A new paired humanization pipeline that optimizes VH and VL simultaneously while monitoring pairing likelihood to prevent structural disruption.
  • De Novo Design: Acting as a scoring layer for generative models (e.g., diffusion models) to ensure generated sequences are biophysically viable and native-like.
• Resource Availability: The authors released the source code, trained models, a user-friendly web server, and the curated datasets (including 21M nanobodies and 3.7M paired sequences) to the community, standardizing nativeness assessment in antibody engineering.

Conclusion

This work represents a significant leap in computational antibody design. By scaling up data, refining the Transformer architecture with focal loss, and introducing a cross-attention mechanism for paired sequences, the authors have created a robust system that not only assesses sequence quality but explicitly predicts the critical compatibility between heavy and light chains, addressing a major bottleneck in the development of next-generation antibody therapeutics.