Adaptive Machine Learning Framework enables Unprecedented Yield and Purity of Adeno-Associated Viral Vectors for Gene Therapy

Abstract

Adeno-associated viral (AAV) vectors for gene therapy are becoming integral to modern medicine, providing therapeutic options for diseases once deemed incurable. Currently, optimizing viral vector purification is a critical bottleneck in the gene therapy industry, impacting product efficacy and safety as well as accessibility and cost to patients. Traditional optimization methods are resource-intensive and often fail to adjust the purification process parameters to maximize the resulting product yield and quality. To address this challenge, we developed a machine learning framework that leverages Bayesian optimization to systematically refine affinity chromatography parameters (sample load, flow rate, and the formulation of chromatographic media) to improve AAV purification. The efficiency of this closed-loop workflow in iteratively optimizing the vector's yield, purity, and transduction efficiency was demonstrated by purifying clinically-relevant serotypes AAV2, AAV5, and AAV9 from HEK293 cell lysates using the affinity adsorbent AAVidity. We show that three cycles of Bayesian optimization elevated yields from a baseline of 70% to 99%, while reducing host-cell impurities by 230-to-400-fold across all serotypes. The optimized parameters consistently produced vectors with high purity and preserved high transduction activity, essential for therapeutic efficacy and safety, demonstrating serotype versatility - a key challenge in AAV manufacturing. By streamlining parameter optimization and enhancing productivity, our adaptive machine learning framework accelerates process development and reduces costs, advancing the accessibility and clinical translation of AAV-based gene therapies.

Keywords: Bayesian optimization; Gene therapy; machine learning; purification; viral vectors.

Figure 1.

Closed-loop optimization of AAV purification by affinity chromatography: (A) selection of an initial random set of process input parameters; (B) experimental data collection; (C) purification performance results recorded in step B are used for model development; (D) build and update a Gaussian process surrogate model using input and output process parameters; (E) implement an acquisition function to select the values of input parameter for the subsequent chromatography test set. The proposed optimization cycle iterates steps (B), (C), (D), and (E) until the optimization objectives (i.e., AAV yield and purity) are reached.

Figure 2.

Computational simulation of the optimization campaign. Sequential acquisition functions qEI, qlogEI, qPI, and qUCB were used to simulate an experimental optimization of total capsids (TC). Simulations showed that the acquisition functions could find optimal parameters in 30 experiments, significantly surpassing a random sampling policy.

Figure 3.

Model-guided optimization of AAV affinity chromatography demonstrates the maximization of AAV yield compared to baseline values obtained without model guidance. Comparison of product yields obtained by purifying AAV2, AAV5, and AAV9 from HEK293 cell lysates using the affinity resin AAVidity operated by machine learning-guided input process parameters.

Figure 4.

(A) Values of capsid yield and purity measured by SEC-HPLC analysis of the elution fractions obtained by purifying AAV2 from HEK293 cell lysates using AAVidity. Process input parameters were selected across three iterations of the closed-loop Bayesian optimization process. (B) Values of capsid titer, HEK293 HCP titer, and HT1080 cell transduction activity in the elution fractions obtained by purifying AAV2 using AAVidity resin operated with process parameters selected in the third iteration of Bayesian optimization.

Figure 5.

(A) Values of capsid yield and purity measured by SEC-HPLC analysis of the elution fractions obtained by purifying AAV9 from HEK293 cell lysates using AAVidity operated with process input parameters selected across three iterations of the closed-loop Bayesian optimization process. (B) Values of capsid titer, HEK293 HCP titer, and HT1080 cell transduction activity in the elution fractions obtained by purifying AAV9 using AAVidity resin operated with process parameters selected in the third iteration of Bayesian optimization.

Figure 6.

(A) Values of capsid yield and purity measured by SEC-HPLC analysis of the elution fractions obtained by purifying AAV5 from HEK293 cell lysates using AAVidity operated with process input parameters selected across three iterations of the closed-loop Bayesian optimization process. (B) Values of capsid titer, HEK293 HCP titer, and HT1080 cell transduction activity in the elution fractions obtained by purifying AAV2 using AAVidity resin operated with process parameters selected in the third iteration of Bayesian optimization.

Figure 7.

SHAP values illustrate the individual parameter contributions to the total capsid yield obtained by purifying AAV2, AAV5, and AAV9 from HEK293 cell lysates using the affinity resin AAVidity. SHAP analysis quantifies the effect of each chromatographic process parameter, supporting data-driven optimization of process conditions. Red dots represent high parameter values, while blue dots indicate low values. The X-axis displays the magnitude and direction of each parameter’s impact on total capsid production.