Optimal trade-off control in machine learning-based library design, with application to adeno-associated virus (AAV) for gene therapy

Abstract

AAVs hold tremendous promise as delivery vectors for clinical gene therapy. Yet the ability to design libraries comprising novel and diverse AAV capsids, while retaining the ability of the library to package DNA payloads, has remained challenging. Deep sequencing technologies allow millions of sequences to be assayed in parallel, enabling large-scale probing of fitness landscapes. Such data can be used to train supervised machine learning (ML) models that predict viral properties from sequence, without mechanistic knowledge. Herein, we leverage such models to rationally trade-off library diversity with packaging capability. In particular, we show a proof-of-principle application of a general approach for ML-guided library design that allows the experimenter to rationally navigate the trade-off between sequence diversity and fitness of the library. Consequently, this approach, instantiated with an AAV capsid library designed for packaging, enables the selection of starting libraries that are more likely to yield success in downstream selections for therapeutics and beyond. We demonstrated this increased success by showing that the designed libraries are able to more easily infect primary human brain tissue. We expect that such ML-guided design of AAV libraries will have broad utility for the development of novel variants for therapeutic applications in the near future.One Sentence SummaryComputational, data-driven re-design of a state-of-the-art therapeutically relevant AAV initial library improves downstream selection for therapeutic uses.

Conformal prediction for the design problem

“…The design problem is a unique setting in which we have control over the data-dependent test input distribution, P X;D , since we choose the procedure used to design an input. In the simplest case, some design procedures sample from a distribution whose form is explicitly chosen, such as an energy-based model whose energy function is proportional to a trained regression model's predictions [10], or whose parameters are set by solving an optimization problem (e.g., to train a generative model) [50,29,12,17,53,70,24,55,74]. In either setting, we know the exact form of the test input distribution, which also absolves the need for density estimation.…”

“…The training input distribution, P X , is also often explicitly known. In protein design problems, for example, training sequences are often generated by introducing random substitutions to a single wild type sequence [12,10,14], by recombining segments of several "parent" sequences [35,52,9,22], or by independently sampling the amino acid at each position from a known distribution [74,67]. Conveniently, we can then compute the weights in Eq.…”

“…Recently, efforts have been made to augment such approaches with machine learning-based strategies; see reviews by Yang et al [72], Sinai and Kelsic [56], Hie and Yang [26], and Wu et al [71] and references therein. For example, one might train a regression model on protein sequences with experimentally measured fitnesses, then use an optimization algorithm or fit a generative model that leverages that regression model to propose promising new proteins [18,12,52,9,69,5,10,36,22,68,74]. Special attention has been given to the single-shot case, where the goal is to design fitter proteins given just a single batch of training data, due to its obvious practical convenience.…”

“…In the experiments presented here, our goal will be as follows. Given training data consisting of protein sequences labeled with experimental measurements of their fitnesses, we will fit a regression model, then sample test sequences (representing designed proteins) according to design procedures used in recent work [10,74] (Fig. 2).…”

“…Sampling designed sequences. Following ideas in [10,74], we design a protein by sampling from sequence distribution whose log-likelihood is proportional to the prediction of the regression model:…”