Student or educator? Sign up with your school email and get 6 months of Pro free.
CoChat

The DOI matches its Crossref record; 68 of the 68 references it cites were found in a registry or an index.

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

Open access

Published version of a preprint

Abstract

Adeno-associated viruses (AAVs) hold tremendous promise as delivery vectors for gene therapies. AAVs have been successfully engineered—for instance, for more efficient and/or cell-specific delivery to numerous tissues—by creating large, diverse starting libraries and selecting for desired properties. However, these starting libraries often contain a high proportion of variants unable to assemble or package their genomes, a prerequisite for any gene delivery goal. Here, we present and showcase a machine learning (ML) method for designing AAV peptide insertion libraries that achieve fivefold higher packaging fitness than the standard NNK library with negligible reduction in diversity. To demonstrate our ML-designed library’s utility for downstream engineering goals, we show that it yields approximately 10-fold more successful variants than the NNK library after selection for infection of human brain tissue, leading to a promising glial-specific variant. Moreover, our design approach can be applied to other types of libraries for AAV and beyond.

Overview

The claims and statistics of this paper haven’t been checked yet.

Free: it reads the open-access text and checks the first 5 claims that cite a source we can check. Create a free account to check every claim.

A few free checks a day · papers already verified open instantly · results are public

Integrity signals

Registry

Crossref · 10.1126/sciadv.adj3786

Cites

68 of 68 references verified0 citation errors · 0 not checkable

Statistics

Figures

Fig. 1. Experimental workflow for generating pre- and post-packaged AAV5–7-mer library data for ML-based library design. Ni, number of reads for each unique insertion sequence i. Experimental data were used to build a supervised regression model where the target variable reflects the packaging success of each insertion sequence. The predictive model was then systematically inverted to design libraries that trace out an optimal trade-off curve between diversity and packaging fitness. Schematic illustration created with BioRender.com.
Fig. 1. Experimental workflow for generating pre- and post-packaged AAV5–7-mer library data for ML-based library design. Ni, number of reads for each unique insertion sequence i. Experimental data were used to build a supervised regression model where the target variable reflects the packaging success of each insertion sequence. The predictive model was then systematically inverted to design libraries that trace out an optimal trade-off curve between diversity and packaging fitness. Schematic illustration created with BioRender.com.CC BY-NC · PubMed Central
Fig. 2. Predictive model architectures and results. (A) Comparison of models for predicting AAV5–7-mer packaging log enrichment scores, using Pearson correlation, and “top-K” Pearson correlation, where K denotes what fraction of top-ranked observed log enrichment test sequences were used. The correlation is between predicted and true log enrichment scores. Seven different models including four neural network (NN) architectures, distinguished by the number of nodes in the hidden layers (100, 200, 500, and 1000). (B) Similar plot to (A) except comparing the use of weighted versus unweighted sequences during training, for the final selected model, NN, 100, and a baseline, linear, pairwise. (C) Schematic illustrations of “linear, IS” (left) and “NN, 4” (right) predictive models. Each model predicts packaging log enrichment from peptide insertion sequence.
Fig. 2. Predictive model architectures and results. (A) Comparison of models for predicting AAV5–7-mer packaging log enrichment scores, using Pearson correlation, and “top-K” Pearson correlation, where K denotes what fraction of top-ranked observed log enrichment test sequences were used. The correlation is between predicted and true log enrichment scores. Seven different models including four neural network (NN) architectures, distinguished by the number of nodes in the hidden layers (100, 200, 500, and 1000). (B) Similar plot to (A) except comparing the use of weighted versus unweighted sequences during training, for the final selected model, NN, 100, and a baseline, linear, pairwise. (C) Schematic illustrations of “linear, IS” (left) and “NN, 4” (right) predictive models. Each model predicts packaging log enrichment from peptide insertion sequence.CC BY-NC · PubMed Central
Fig. 3. Experimental titers versus predicted log enrichment scores. The five variants were selected to span a broad range of predicted log enrichment scores. Log enrichment scales are computed using natural logarithm. Experimental titers are measured on three biological replicates.
Fig. 3. Experimental titers versus predicted log enrichment scores. The five variants were selected to span a broad range of predicted log enrichment scores. Log enrichment scales are computed using natural logarithm. Experimental titers are measured on three biological replicates.CC BY-NC · PubMed Central
Fig. 4. Designed AAV5-based 7-mer insertion libraries. Each point in (A) represents a theoretical library designed with our diversity-constrained optimal library approach, with one particular diversity constraint, λ (higher values yields more diverse libraries). Entropy indicates diversity of the library distribution, while MPLE indicates overall library fitness; both quantities were computed from the theoretical library distribution. The baseline NNK library is denoted with a black “×”, while a cyan “×” denotes the filtered uniform library that is uniform over all 21-mer nucleotide sequences except for those containing stop codons. Three designed libraries have been circled and labeled D1 to D3 for reference. Because of the non-convex optimization problem, some dots are suboptimal (i.e., lie strictly below or to the left of other dots) and are therefore further from the optimal frontier but are displayed for completeness. (B to D) Designed library parameters (probability of each amino acid at each position) for the three designed libraries D1 to D3, respectively, highlighted in (A).
Fig. 4. Designed AAV5-based 7-mer insertion libraries. Each point in (A) represents a theoretical library designed with our diversity-constrained optimal library approach, with one particular diversity constraint, λ (higher values yields more diverse libraries). Entropy indicates diversity of the library distribution, while MPLE indicates overall library fitness; both quantities were computed from the theoretical library distribution. The baseline NNK library is denoted with a black “×”, while a cyan “×” denotes the filtered uniform library that is uniform over all 21-mer nucleotide sequences except for those containing stop codons. Three designed libraries have been circled and labeled D1 to D3 for reference. Because of the non-convex optimization problem, some dots are suboptimal (i.e., lie strictly below or to the left of other dots) and are therefore further from the optimal frontier but are displayed for completeness. (B to D) Designed library parameters (probability of each amino acid at each position) for the three designed libraries D1 to D3, respectively, highlighted in (A).CC BY-NC · PubMed Central
Fig. 5. Comparison of ML-designed libraries D2 and D3 to the NNK library. (A) Experimental titers (viral genome/ml) plotted against the MPLE. ***P < 0.001 compared to NNK. (B) Comparison of the effective number of variants present in each library after packaging and post-packaged libraries are labeled with the “-post” suffix. (C) Experimental titers and effective number of variants for D2, D3, and NNK DNA libraries (pre-packaging selection), and the NNK-post library (post-packaging selection). The NNK-post library represents the NNK library after one round of packaging selection. One-way ANOVA followed by Tukey test (**P < 0.01 compared to D2). In all cases, experimental titers are measured on three biological replicates. Graphs show means ± SD.
Fig. 5. Comparison of ML-designed libraries D2 and D3 to the NNK library. (A) Experimental titers (viral genome/ml) plotted against the MPLE. ***P < 0.001 compared to NNK. (B) Comparison of the effective number of variants present in each library after packaging and post-packaged libraries are labeled with the “-post” suffix. (C) Experimental titers and effective number of variants for D2, D3, and NNK DNA libraries (pre-packaging selection), and the NNK-post library (post-packaging selection). The NNK-post library represents the NNK library after one round of packaging selection. One-way ANOVA followed by Tukey test (**P < 0.01 compared to D2). In all cases, experimental titers are measured on three biological replicates. Graphs show means ± SD.CC BY-NC · PubMed Central
Fig. 6. ML-designed AAV library for primary brain tissue infection. (A) General workflow of the primary adult brain infection study. (B) Effective number of variants (calculated from entropy) in NNK post-brain infection versus D2 post-brain infection; D2 post-brain infection exhibits a ~10-fold increase in effective number of variants compared to that of NNK post-brain infection. (C) Empirical probabilities of each amino acid at each position for D2 post-packaging and post-brain infection. (D) Scatterplots illustrating the behavior of individual variants over packaging and primary brain selection. Each axis shows the (log) prevalence of the variant in each library, as a fraction of reads in the library. For each library, variants in the top 20% are determined by first sorting unique variants by read count in descending order and then counting the number of unique variants comprising 20% of the total sequencing reads. Variants in the top 20% after packaging are colored blue, while those in the top 20% after brain selection are colored yellow. Those variants in the top 20% of both packaging and selection are colored green. The annotated colored numbers indicate the number of variants of each colored pool. A pseudo-count of 1 was added to each variant in each library before plotting. See fig. S11 for additional versions of (D) displaying variants in the top 50 and 80% of each library. (E) Cell-specific AAV validation (VVKQRGD) selected from the post-brain infection pool [green, glial fibrillary acidic protein (GFAP) marker; red, AAV infected cells; scale bars, 100 μm; CP, cortical plate; IZ, intermediate zone; VZ, ventricular zone; DAPI, 4′,6-diamidino-2-phenylindole). Schematic illustration created with BioRender.com.
Fig. 6. ML-designed AAV library for primary brain tissue infection. (A) General workflow of the primary adult brain infection study. (B) Effective number of variants (calculated from entropy) in NNK post-brain infection versus D2 post-brain infection; D2 post-brain infection exhibits a ~10-fold increase in effective number of variants compared to that of NNK post-brain infection. (C) Empirical probabilities of each amino acid at each position for D2 post-packaging and post-brain infection. (D) Scatterplots illustrating the behavior of individual variants over packaging and primary brain selection. Each axis shows the (log) prevalence of the variant in each library, as a fraction of reads in the library. For each library, variants in the top 20% are determined by first sorting unique variants by read count in descending order and then counting the number of unique variants comprising 20% of the total sequencing reads. Variants in the top 20% after packaging are colored blue, while those in the top 20% after brain selection are colored yellow. Those variants in the top 20% of both packaging and selection are colored green. The annotated colored numbers indicate the number of variants of each colored pool. A pseudo-count of 1 was added to each variant in each library before plotting. See fig. S11 for additional versions of (D) displaying variants in the top 50 and 80% of each library. (E) Cell-specific AAV validation (VVKQRGD) selected from the post-brain infection pool [green, glial fibrillary acidic protein (GFAP) marker; red, AAV infected cells; scale bars, 100 μm; CP, cortical plate; IZ, intermediate zone; VZ, ventricular zone; DAPI, 4′,6-diamidino-2-phenylindole). Schematic illustration created with BioRender.com.CC BY-NC · PubMed Central

Verification

Claims the paper makes

The claims and statistics of this paper haven’t been checked yet.

Free: it reads the open-access text and checks the first 5 claims that cite a source we can check. Create a free account to check every claim.

A few free checks a day · papers already verified open instantly · results are public

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

Self-check

Percentages, sums and reported differences re-computed from the paper’s own numbers.

The self-check needs the paper’s full text.

Citations

Cites

Show all 65 references

Research works printed but not linked: 1

  1. K. P. Murphy Machine Learning: A Probabilistic Perspective (MIT Press 2012).

Cited by

See all 64

Metadata

Authors
Zhu, D., Brookes, D. H., Busia, A., Carneiro, A., Fannjiang, C., Popova, G., Shin, D., Donohue, K. C., Lin, L. F., Miller, Z. M., Williams, E. R., Chang, E. F., Nowakowski, T. J., Listgarten, J., & Schaffer, D. V.
Affiliations
  • QB3 (US)
  • University of California, Berkeley (US)
  • University of California, San Francisco (US)
Corresponding author
  • Zhu, Danqing
  • Brookes, David H.
  • Busia, Akosua
  • Listgarten, Jennifer
  • Schaffer, David V.
Venue
Science Advances
Published
2024-01-26
Volume
10
Issue
4
Pages
eadj3786-eadj3786
ISSN
2375-2548 (ISSN-L)
OpenAlex
W4391203107
Cited by
63
References
65
Citations by year
2026: 14 · 2025: 21 · 2024: 18
FWCI
14.02
Type
journal article
Language
en
Open access
Licence
cc-by-nc
Locations
  • Science Advances (publishedVersion, cc-by-nc)
  • bioRxiv (Cold Spring Harbor Laboratory) (acceptedVersion, cc-by-nc-nd)
  • PubMed Central (publishedVersion, cc-by-nc)
  • PubMed (publishedVersion)
  • eScholarship (California Digital Library) (submittedVersion, cc-by-nc)
  • Rare & Special e-Zone (The Hong Kong University of Science and Technology) (submittedVersion)
Funders
  • National Science Foundation (1752814, 2146752, DGE 2146 752)
  • Shurl and Kay Curci Foundation
  • National Institutes of Health (5r01gm139338-04, 1752814, 5R01GM139338)
  • National Institute of Mental Health (UF1MH130700, F32MH118785)
  • National Institute of General Medical Sciences (R01 GM139338)
  • National Institute of Neurological Disorders and Stroke (R01NS126397)
Topics
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Genetics › Virus-based gene therapy research
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Molecular Biology › RNA Interference and Gene Delivery
  • Life Sciences › Biochemistry, Genetics and Molecular Biology › Molecular Biology › CRISPR and Genetic Engineering
Keywords
adeno-associated virus, library design, machine learning, gene therapy vectors, human brain tissue, viral genome assembly
MeSH terms
Machine Learning, Dependovirus, Brain, Humans, Genetic Therapy, Peptide Library

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: Optimal trade-off control in machine learning–based library design, with application to adeno-associated virus (AAV) for gene therapy. https://app.cochat.ai/papers/10.1126/sciadv.adj3786
The paper
Zhu, D., Brookes, D. H., Busia, A., Carneiro, A., Fannjiang, C., Popova, G., Shin, D., Donohue, K. C., Lin, L. F., Miller, Z. M., Williams, E. R., Chang, E. F., Nowakowski, T. J., Listgarten, J., & Schaffer, D. V. (2024). Optimal trade-off control in machine learning–based library design, with application to adeno-associated virus (AAV) for gene therapy. Science Advances. https://doi.org/10.1126/sciadv.adj3786

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.