Table 4 List of studies associated with improved synthetic Ab library designs based on deep sequencing big data assessments

Maruthachalam et al. [36]

Deep sequencing of synthetic Ab libraries helped produce diversified sub-libraries with fine-tuned CDR lengths for improved target antigen recognition

▪ Uses a single-framework synthetic Ab library

▪ The only diversified CDRs include H1–H3 and L3

▪ Target antigen dependent

Chen et al. [39]

The authors generated a synthetic Ab library against enterovirus antigens. Deep sequencing analysis reveals that heavy chains of the enterovirus-specific Abs are conserved

▪ The Ab library is based on peripheral blood samples of enterovirus-infected donors

Chen et al. [40]

The deep sequencing of immunized mice enabled the reverse-engineering of an Ab response while helping guide the construction of a synthetic Ab library containing natural features

▪ The synthetic library complexity is limited to natural Ab repertoires

Larman et al. [41]

The successful development of a rationally designed synthetic Ab library was used in downstream applications involving deep sequencing-associated Ab discovery

▪ Limited to a ribosome-display format

▪ The only diversified CDRs include H2–H3 and L3

Tiller et al. [42]

Ravn et al. [43]

Frigotto et al. [44]

These studies demonstrate the effective large-scale, quality controlled validations of engineered synthetic Ab libraries by extensive deep sequencing analysis

▪ The longer reads show reduced data quality due to greater CDR complexities and read-out errors

Li et al. [48]

A machine learning model, trained on extensive deep sequencing datasets, was able to effectively engineer various synthetic Ab libraries in silico. These are highly diverse and target-specific

▪ Requires target binding data for supervised training

▪ Demands supervised fine-tuning of pretrained language models

Shuai et al. [49]

This study presents generative language model employs bidirectional context for designing Ab sequence spans of varying lengths. It led to the effective in silico design of synthetic Ab libraries containing desirable biophysical features

▪ Requires data training from natural Ab sequences for targeted infilling of residue spans and full-length sequence generation

▪ Desirable biophysical features require the generation of large-scale full-length Ab sequences for model training

Amimeur et al. [50]

A machine learning model, based on a generative adversarial network, effectively generates feature-controlled synthetic Ab libraries. The method included transfer learning; thus enabling chemical and biophysical biases

▪ The proof-of-concept validation library only contains 100 k sequences

▪ Validation of the library failed to include antigen selections and identify target selective Abs

Shin et al. [51]

An autoregressive generative machine learning model was able to predict functional Ab sequences for the engineering of synthetic nanobody libraries. These possessed high levels of Ab expression, solubility, and stability

▪ Restricted to nanobody libraries

▪ Based on a small library that requires further affinity maturations to identify strong Ab binders