Secure Biosystems Design Scientific Focus Area

Extending the DropSynth method for pooled synthesis of gene entanglement libraries and parallel testing in cells. Multiplex reactions across hundreds of thousands of emulsion droplets enable simultaneously assembling tens of thousands of genes, which can be released from the droplets and subsequently amplified by PCR or cloned into shuttle vectors en masse.

Synthesizing genes in a cheap, fast, and parallel fashion enables large-scale genetic design testing that is essential for various synthetic biology and biocontainment applications. Commercial oligonucleotide synthesis manufacturers deliver tens of thousands of short oligonucleotides (for example, 300 nucleotides in length) that can only test very short sequence designs. However, our work requires synthesis capabilities that generate genes with an average length of 1000 nucleotides, which has remained challenging and expensive.

The DropSynth method leverages picoliter emulsion-based gene synthesis to produce gene libraries with thousands of designs of up to 700 base pairs at prices cheaper than commercial gene synthesis. We use this method to produce libraries of hundreds to thousands of gene-sized sequences to test the sequence designs from our computational design pipelines for sequence entanglement studies. We are also developing compatible protocols that leverage DropSynth for a variety of downstream high-throughput screening applications. Our team is generating hundreds to thousands of entangled sequences to be tested simultaneously through homologous recombination en masse in our target organism of interest.

Our project requires gene-editing tools that enable stable, efficient, and rapid genome modification to enhance native capabilities, introduce new capabilities, and implement biocontainment approaches. We will develop genome engineering tools for the environmental Pseudomonas species, including classic (transposon mutagenesis and gene removal/replacement using homologous recombination) and modern (recombineering, CRISPRi, and CRISPR/Cas) approaches. CRISPR-based gene-editing tools have been developed for P. aeruginosa and P. putida, but have yet to be successfully applied in P. fluorescens.

Our team’s experience places us at the forefront of gene editing and circuit design in environmental microorganisms, which is critical for realizing our layered containment approach. Co-principal investigator Jeffrey Gralnick’s laboratory developed recombineering and CRISPR-Cas counter-selection tools in Shewanella oneidensis, a close relative of environmental Pseudomonads. In addition, the LLNL team has extensive experience in genome editing in a diverse range of Gram-negative bacteria, including E. coli, Caulobacter crescentus, Rhodopseudomonas palustris, and Ralstonia eutropha.

Identifying regulatory elements, such as promoters and ribosome binding sites (RBS), that function in a broad or host-specific manner enables the design of gene regulatory circuits that produce species-/genus-specific outputs, thereby providing an additional framework for safeguarding against horizontal gene transfer. We developed in vivo and in vitro platforms that enable more than ten thousand metagenomically mined regulatory sequences to be quantitatively assayed for transcriptional and translational activity in multiplexed format in a diverse range of microorganisms.

For in vivo assays, regulatory sequences from annotated genomes are synthesized as an oligo library that includes unique barcodes, flanking restriction sites, and amplification sequences. The oligo library is amplified and cloned as a pool into species-specific vectors for transformation into recipients. RNA-seq, DNA-seq, and FACS-seq enable us to quantify transcription and translation activity. This approach is useful in biocontainment, as is evidenced by the finding of differences in the capacity of B. subtilis, E. coli, and P. aeruginosa to utilize exogenous regulatory sequences. This knowledge provides a proof-of-concept for circuit design with species-specific output patterns.

To extend the regulatory element characterization to microbes without an established transformation protocol, we developed a high-throughput cell-free framework—the DNA regulatory element analysis by cell-free transcription and sequencing (DRAFTS)—that produces transcription measurements highly correlated with those from the in vivo platform. DRAFTS enabled thousands of regulatory elements to be characterized across ten diverse species relevant to biotechnology and the gut microbiome. Leveraging LLNL’s extensive expertise in cell-free systems and in vivo screening, we are extending the regulatory element characterization to a range of Pseudomonads under conditions relevant to soils. Our goal is to identify promoters and RBS that function robustly in narrow or broad host ranges.

FAST is an LLNL-developed, open source, deep learning software toolkit for learning experimentally derived protein-ligand binding affinity values from 3D structures. (SG=spatial graph, CNN=convolutional neural network. Nodes in the spatial graph are heavy atoms.) We will extend this modelling approach to learn 3D-protein structure configurations associated with the positive outcomes of successfully engineered sequences for entanglement.

We leveraged LLNL’s experience in structural and computational biology to augment the Constraining Adaptive Mutations using Engineered Overlapping Sequences (CAMEOS) algorithm with protein structural constraints and design specifications obtained from available crystal structures. However, representing 3D structural features for modeling can be challenging given the complexity of protein structures. A potential unique value of deep learning is to simultaneously learn the relevant features and the experimentally derived response function.

The LLNL-developed, open source, deep learning software toolkit called fusion models for atomic and molecular structures (FAST) learns experimentally derived binding affinity values from the 3D atomistic representation of a protein-small molecule interaction. We will extend this modelling approach to learn 3D protein-structure configurations associated with the positive outcomes of successfully engineered entangled sequences. Input is the 3D homology model of the essential or functional protein and the algorithm encodes a spatial graph to capture pairwise interactions and a 3D voxel cube to capture additional spatial relationships. The model learns a compressed protein representation to predict viable or non-viable protein configurations.

A platform that builds on LLNL’s recent advances in machine learning, design optimization, simulations, and experimental assays to support LLNL’s high-performance computing (HPC) mission: building new methods for applying HPC to solve scientific problems using state-of-the-art machine learning algorithms. We will use this platform to develop an iterative active-learning framework for sequence entanglement design.

LLNL established a framework for applying HPC to solve scientific problems using machine learning (ML) algorithms with feedback from wet-lab experiments and computational simulations. We demonstrated an example of this framework with a drug discovery pipeline, which searches through chemical space to find molecules that satisfy the design criteria for anti-cancer therapies and then evaluates candidates using property-prediction models. The loop represents the search process as an iterative search through chemical space, with break points to collect feedback from physical- or simulation-based experiments. One of the innovations of this approach is to evaluate whether a more efficient search can be done when chemical space is converted into a continuous feature space that is amenable to gradient-based optimization algorithms.

Collecting data on new molecules is a critical component, with a ML loop guiding the collection of new experimental feedback and validation for new molecules that fall outside of the domain of applicability of the initial training data. LLNL’s HPC clusters support every aspect of this platform, including the extensive hyperparameter search for deep learning model training, the training of autoencoders for the Generative Molecular Design loop on tens to hundreds of millions of compounds, and the iterative multi-property prediction design search—which can scale to search the design space in parallel on a large cluster.

We are developing an iterative active-learning pipeline by which ML models acquire and guide new experimental data on entanglement. By iteratively including newly generated data (fitness, expression), we augment and refine the generative models that form the basis of CAMEOS to increase the success of double-encoding entanglement designs. Additionally, these methods can be adapted to work without the acquisition of new data, enabling in silico screening of designed variants, thereby further improving the likelihood of synthesizing functional entangled sequences as experimental validation data is collected.

Leveraging the extensive expertise of our co-principal investigators William Bentley and Gregory Payne, we will employ and advance microfluidic platforms to meet various conditions and needs in evaluating the effectiveness of our containment circuits. These platforms provide a quantitative means to study spatially and temporally heterogenous environments that are characteristic of soils and the rhizosphere. Coupled with real-time, high-resolution microscopy, we will be able to examine how steady-state, non-steady state, and fluctuating concentration gradients of environmental variables affect containment in a spatially resolved fashion.

The use of transparent soils, multiple hydrogel layers, and 2D fluidic systems, will allow us to further examine how different spatial arrangements influence containment. We can track cell migration and quantify swimming parameters using simple CCD camera-based approaches developed by our team. Furthermore, our microfluidic devices can modulate redox context and detect electrochemical signals that can be correlated with synthetic biology inputs and outputs as a novel non-optical imaging modality to probe sensing and containment dynamics in real time.