Evaluation of Molecular Simulations and Deep Learning Prediction of Antibodies’ Recognition of TRBC1 and TRBC2

Xincheng Zeng, Tianqing Wang, Yue Kang, Ganggang Bai, Buyong MaAntibodies 2023 · DOI

TL;DR

This paper introduces a novel AI model, "Digiwiser," that leverages deep learning and molecular simulations to predict how antibody mutations affect their binding affinity to closely related protein targets, TRBC1 and TRBC2. This work is crucial for developing highly specific antibody therapies that can precisely target cancer cells while sparing healthy ones.

What Problem Does This Paper Address?

The development of targeted antibody therapies for diseases like cancer often hinges on the ability to distinguish between highly similar protein targets. In this case, the focus is on TRBC1 and TRBC2, two nearly identical proteins that are part of T cell receptors. While antibodies can be engineered to target these proteins, achieving high specificity is challenging due to their structural similarity. This similarity means that an antibody designed to bind TRBC1 might also bind TRBC2, leading to unintended side effects. The paper aims to address this challenge by developing computational tools that can accurately predict how specific mutations in an antibody will alter its binding preference between TRBC1 and TRBC2. This capability is vital for rational antibody design and engineering, enabling the creation of therapeutics that can selectively eliminate malignant T cells expressing one TRBC variant while leaving T cells expressing the other unharmed.

Methods

The study employs a two-pronged approach combining artificial intelligence (AI) and molecular simulations.

  • AI Model for Affinity Prediction: The core AI component is a "listwise ranking model" called "Digiwiser." This model is trained on existing datasets (SKEMPI and AB-Bind) containing antibody-antigen complexes with known mutations and their corresponding binding free energy changes (ΔΔG). Instead of predicting absolute ΔΔG values, the model predicts a "relevance score" that ranks the binding affinity changes. This approach is designed to be more robust to noise and better capture the relative impact of mutations. Pre-trained language models (AntiBERTY and ProtBert) are used to generate embeddings for antibody and antigen sequences, which are then fed into transformer encoder layers.
  • Molecular Dynamics (MD) Simulations: Long-time MD simulations (up to 1000 ns) are used to observe the dynamic interactions between antibodies and their targets. This helps to understand how stable or unstable antibody-antigen complexes are over time and to reveal the underlying mechanisms of recognition.
  • Free Energy Perturbation (FEP) Calculations: FEP is used to quantitatively assess the binding affinity changes caused by specific mutations. This method, while computationally intensive, provides a more rigorous calculation of free energy changes compared to simpler methods. The study uses FEP to verify the AI model's predictions and to gain deeper insights into the energetic contributions of mutations.

The novelty lies in the combination of a listwise ranking AI model with extensive MD and FEP simulations to specifically address the challenge of distinguishing between highly homologous targets like TRBC1 and TRBC2.

Datasets

  • Training Data: The AI model was trained on the SKEMPI [27] and Antibody–Bind (AB-Bind) [28] datasets. These datasets comprise approximately 1413 antibody-antigen pairs with single-site and multi-site mutations and their associated binding free energy change values (ΔΔG).
  • Evaluation Data: The study uses experimental data for the JOVI-1 antibody and its variants binding to TRBC1 and TRBC2. Table 2 lists 18 simulated antibody-antigen complexes, including wild-type and mutant versions, along with their experimental affinities and simulated stable times.

Key Insights and Ideas

  • AI Model's Strength in Ranking: The Digiwiser AI model shows promise in ranking the effects of antibody mutations on binding affinity, outperforming traditional in silico methods in correlation with experimental data. This is particularly useful for identifying trends in affinity changes.
  • Challenge of Distinguishing Homologs: The AI model, while good at ranking, struggles to precisely distinguish between the subtle differences in binding affinity for TRBC1 versus TRBC2. This highlights a significant challenge in developing AI models for highly homologous targets.
  • MD Reveals Dynamic Recognition: MD simulations reveal that antibody-antigen interactions are highly dynamic. Hydrogen bond networks fluctuate and new interactions can form during simulations, offering insights beyond static crystal structures.
  • Entropy Plays a Role: The study suggests that the entropy of the antibody-antigen complex influences binding stability. Weakly bound complexes might maintain higher antibody entropy, which can be transferred to the antigen in strongly bound complexes.
  • Specific Residues are Key: The T28 residue in the JOVI-1 antibody and its interaction with K120 in TRBC1 are identified as crucial for selective binding. Disrupting this interaction can significantly alter the antibody's preference.

Key Results

  • AI Model Performance: The Digiwiser model achieved Pearson correlations of 0.74-0.89 and Spearman correlations of 0.72-0.78 on validation datasets, outperforming other methods. For the specific TRBC1/2 targets, the model showed better Pearson correlation for TRBC1 (0.543) but superior Spearman ranking correlation for TRBC2 (0.899).
  • MD Simulation Stability: Three out of eighteen simulated complexes with the lowest binding affinities (JOVI-1 TRBC2, mut1b JOVI-1/T28 TRBC2, and mut3a JOVI-1/KFN TRBC1) dissociated during the 1000 ns MD simulations, consistent with their lower experimental affinities.
  • FEP Accuracy: FEP calculations generally reproduced experimental binding affinity changes, with average errors of 1.28 kcal/mol for TRBC1 and 2.09 kcal/mol for TRBC2. The higher error for TRBC2 suggests that JOVI-1 binding to TRBC2 is more sensitive to mutations.

Limitations and Future Work

  • AI Model Sensitivity to Antigen Differences: The AI model's performance is better when antibodies bind to identical antigens. It struggles to capture small variations between highly homologous antigens like TRBC1 and TRBC2, leading to similar prediction scores for both. The authors suggest including more diverse antigens in future training datasets.
  • Limited Experimental Data: The availability of experimental data for antibody engineering is often limited, which can lead to overfitting in AI models.
  • FEP Sampling Time: The FEP simulations used a 1 ns sampling time window. Longer simulation times could potentially improve accuracy, especially for TRBC2.
  • Future Directions: The authors emphasize the need to combine AI with molecular simulations for a comprehensive understanding of antibody-antigen interactions, particularly for systems with high dynamic contributions. They also suggest developing more accurate FEP protocols and incorporating antigens with high sequence identity into training sets to improve model robustness and applicability. The development of protein language models is also highlighted as a promising avenue for improving antibody activity and properties.

Summary updated

Abstract

T cell receptor β-chain constant (TRBC) is a promising class of cancer targets consisting of two highly homologous proteins, TRBC1 and TRBC2. Developing targeted antibody therapeutics against TRBC1 or TRBC2 is expected to eradicate the malignant T cells and preserve half of the normal T cells. Recently, several antibody engineering strategies have been used to modulate the TRBC1 and TRBC2 specificity of antibodies. Here, we used molecular simulation and artificial intelligence methods to quantify the affinity difference in antibodies with various mutations for TRBC1 and TRBC2. The affinity of the existing mutants was verified by FEP calculations aided by the AI. We also performed long-time molecular dynamics simulations to reveal the dynamical antigen recognition mechanisms of the TRBC antibodies.
BibTeX
@inproceedings{zeng2023evaluation,
  author = {Xincheng Zeng and Tianqing Wang and Yue Kang and Ganggang Bai and Buyong Ma},
  title = {{Evaluation of Molecular Simulations and Deep Learning Prediction of Antibodies’ Recognition of TRBC1 and TRBC2}},
  booktitle = {Antibodies},
  year = {2023},
  doi = {10.3390/antib12030058},
}
Figure