In silico design of immunogenic antigen cocktail via affinity maturation-guided optimization

Abstract

Summary: The increasing emergence of new virus strains with increased infectiousness necessitates a more proactive approach for effective vaccine design. To achieve this goal, it is critical to shift the vaccine design paradigm from traditional approaches that rely on expert intuition and experimental methods toward data-driven strategies that leverage in silico design and virtual screening. In this work, we propose a computational pipeline for designing an optimized immunogenic cocktail that can boost the immune response. The proposed pipeline consists of two stages, where potential antigen candidates are identified in the first stage, followed by the optimal selection and combination of the candidates in the second stage to maximize the expected immunogenicity. We leverage predictive models trained using deep mutational scanning data to drive the candidate antigen selection process based on three selection criteria-namely, binding affinity between viral protein and receptor, antibody escape probability, and sequence diversity. To identify the optimal cocktail within the pool of selected antigens, we adopt a combinatorial optimization framework, where the cocktail design is iteratively refined based on the expected efficacy predicted by a sequence-based computational model of affinity maturation. Validation of the designed cocktails through structure-based affinity maturation simulation demonstrates the efficacy of the proposed modular framework for designing an optimized immunogenic cocktail.

Availability and implementation: The code for cocktail design is available in https://github.com/nafizabeer/Antigen_Cocktail_Design.

Figure 1.

In silico design and validation of antigen cocktail. The framework consists of three components: candidate antigen generation, cocktail design, and validation. Candidate antigen generation—the candidate Ags are generated based on ACE2-RBD binding, Ab-escape probabilities, and sequence diversity. The first two criteria are dictated by predictor networks trained with DMS data while the latter is achieved implicitly by clustering of antigen sequence embeddings from a pLM. Cocktail design—the iterative optimization of cocktail is guided by efficacy feedback from a sequence-based AM model. Cocktail validation—the immunogenic response of designed cocktails are validated via more complex computational model of AM which utilizes the structural information of antigens.

Figure 2.

Sequence diversity of 20 candidate antigens. Histogram is generated based on the Hamming distances among unique pairs (190 pairs) of 20 selected antigens for cocktail optimization.

Figure 3.

Optimization traces for affinity maturation guided cocktail design. For Bayesian optimization (a) and genetic algorithm (b, c), traces for five trials show the best fitness (efficacy) value found at each iteration. The fitness at 0^th iteration corresponds to the best cocktail out of five randomly sampled cocktails (used to initialize the optimization strategy). We also show the antigens present in the best cocktail at the 0^th and final iteration. For example, 4-5-18 denotes the antigen IDs of the best cocktail at the corresponding iteration.

Figure 4.

Germinal center dynamics for three different cocktails. The first two rows show cases for top two cocktails and the last row corresponds to the worst cocktail found during the combinatorial optimization search in Section 3.3. For every cocktail, the affinity maturation simulation over 21 days is repeated 5 times, and the average value of the quantity of interest is plotted with a margin of one standard deviation. The left column shows the progress of B-cell population size in the germinal center. The maturation of the B-cells toward the antigens in cocktail is shown in the middle column as an average value of affinities over all centrocytes and centroblasts in the germinal center. The count of unique antibody sequences (the right column) over the simulation period illustrates the sequence diversity in the germinal center.