Neural Network Made with Proteins: Index #58
Michael Elowitz & Co. have done it again.
Cells are bombarded with signals, shifting environments, cagey neighbors. And through all of this noise, through all of this drama, they must make the right decision, or die. Synthetic biologists want to engineer similar sorts of decision-making systems to build ‘smarter’ therapies that, say, could differentiate a cancer cell from a non-cancer cell.
One of the key ways that biology makes decisions is via winner-take-all neural networks. In this type of network, if there are 10 neurons in a layer, only the neuron with the most inputs will fire. The rest of the neurons are silent.
Three things are needed to build a winner-take-all network:
Signal classification; input X leads to output X, input Y leads to output Y.
Outputs are all-or-none; there’s no intermediate state.
The decision threshold — this or that — can be tuned.
Synthetic classification systems have already been built, many times, in test tubes and living cells. This is nothing new. In the last eleven years:
A neural network was built via DNA strand displacement in a test tube (2011).
Winner-take-all networks were built with DNA for digit classification (2018).
mRNA-based classifiers in HeLa cells discriminate gene expression levels (2011)
But building another version of a winner-take-all neural network, at the level of proteins, is a “hot damn”-level feat. Still, Zibo Chen, James Linton, Rongui Zhu and Michael Elowitz have done it. (Aside: The first two papers, above, are from Erik Winfree and Lulu Qian, both of whom work about three minutes from Elowitz’s door.)
Before this preprint, synthetic classifiers were built mainly from mRNA and DNA. Nucleic acids are easy to program, but they are functionally distant from a cellular output. Proteins are the ideal tool for building a synthetic classifier because a protein can waltz into the nucleus, latch onto DNA, and immediately trigger a genetic response.
OK, that’s the introduction. Now for the fun part.
How It Works
Elowitz’s team used three protein parts to build the neural network:
A collection of custom-made heterodimer proteins. These are little proteins that selectively recognize, and link up with, a partner.
Split protease enzymes. These are proteins that chew up peptide bonds. The Caltech team used two types of proteases, taken from viruses that infect tobacco plants, and split each of them in half. When the two halves come together, the completed protease recognizes, and cleaves, a unique peptide sequence.
A degron, or short protein snippet that accelerates protein degradation.
These three proteins all come together to build a working neural network. The custom-made heterodimers (there are two sets, X and Y) are auto-inhibiting; that is, they bind to one another. Their default mode is ‘OFF’. These paired heterodimers are fused to one-half of a protease enzyme. This split enzyme is, in turn, fused to a degron.
When a heterodimer input is added to a cell (specifically, HEK239T kidney cells), the auto-inhibiting heterodimers are disrupted. They unlatch and open up. These “open” heterodimers (fused to the half-protease enzyme and degron) are then free to travel through the cell until they find and bind with their partner (a heterodimer that carries the other half of the protease enzyme).
Thus united, a completed protease will do one of two things: Either it will self-activate by cleaving off its destabilizing degron (switch ‘ON’), or it will bump into another protease. When two proteases come face-to-face, each inactivates the other by cleaving off its dimerization domain; the protein fusions switch ‘OFF’. The weights of this neural network are tuned by adding in more or less of an input signal.
Designs are great and all, but this is biology. Experiments reign supreme.
Fortunately, these scientists are really smart. Before sitting down at the bench and spending the next 17 years trying to build this thing, they inhaled, exhaled, and charted a rational course: model everything on a computer.
What a nightmare that must have been!
From the preprint (bold, my own):
Modeling these reactions required a set of 158 ordinary differential equations containing 1238 terms. This complexity, which exceeds that of most previous synthetic biological circuits, provokes the question of whether the system could actually function in mammalian cells.
Alas, the thing actually worked in mammalian cells. Each piece of the protein jigsaw was expressed on an individual plasmid. Those plasmids were jammed into cells. And then, miraculously — like a calculus student’s nightmare come to life — the cells classified varied input signals.
The authors have named this creation a perceptein network, an homage to the perceptron classification algorithm that was first devised in 1943.
A Little Takeaway
If you’re still reading this, think for a moment about the repressilator, toggle switch, or any other version 1.0 genetic circuit. It has been two decades since those landmark papers in synthetic biology, and it feels as if we’ve come full circle. Did anyone expect that the same lab that pioneered early genetic circuits — by taking three genes, putting them into bacteria, and watching them glow beneath a microscope — would today write 158 ordinary differential equations and build a winner-take-all neural network, from proteins, in living cells? Einstein initiated a radically new way of thinking in physics, but his leadership was supplanted, exactly 20 years later, by the cadre that emerged around quantum mechanics — Heisenberg, Schrödinger, Dirac. Spearheading innovation in any field, after two decades, seems a relatively rare feat.
If the 21st century is the era of biology, then I’m excited by the pace of our achievements; particularly the unification of computational and experimental work. We are only two decades in, and the pace of progress is accelerating. If the leap from repressilator to winner-take-all neural network were marked by a single stride, then I suspect that an even greater leap in circuit complexity will be published in about five years’ time. Mark your calendar and call me on my B.S.
Read more at bioRxiv.
(↑ = recommended article, * = open access, † = review article )
Assembly, Synthesis & Sequencing
*De novo assembly of human genome at single-cell levels. Xie H…Tang F. Nucleic Acids Research. Link
*Targeted de novo phasing and long-range assembly by template mutagenesis. Li S…Levy D. Nucleic Acids Research. Link
↑*Comparing the accuracy and efficiency of third generation DNA barcode sequencing: Oxford Nanopore Technologies versus Pacific Biosciences. Cuber P…Misra R. bioRxiv (preprint). Link
↑Multimodal perception links cellular state to decision making in single cells. Kramer BA…Pelkmans L. Science. Link
*Microbial genomic trait evolution is dominated by frequent and rare pulsed evolution. Gao Y & Wu M. Science Advances. Link
*Auxotrophic and prototrophic conditional genetic networks reveal the rewiring of transcription factors in Escherichia coli. Gagarinova A…Babu M. Nature Communications. Link
*Deciphering polymorphism in 61,157 Escherichia coli genomes via epistatic sequence landscapes. Vigué L…Weigt M. Nature Communications. Link
Biomanufacturing & Metabolic Engineering
Rescuing yeast from cell death enables overproduction of fatty acids from sole methanol. Gao J…Zhou YJ. Nature Metabolism. Link
↑Engineering site-selective incorporation of fluorine into polyketides. Sirirungruang S…Chang MCY. Nature Chemical Biology. Link
Methanol biotransformation toward high-level production of fatty acid derivatives by engineering the industrial yeast Pichia pastoris. Cai P…Zhou YJ. PNAS. Link
*Overexpression of the scopoletin biosynthetic pathway enhances lignocellulosic biomass processing. Hoengenaert L…Vanholme R. Science Advances. Link
Data-driven and model-guided systematic framework for media development in CHO cell culture. Hong JK…Lee D-Y. Metabolic Engineering. Link
*Single cell mutant selection for metabolic engineering of actinomycetes. Akhgari A…Metsä-Ketelä M. Metabolic Engineering. Link
Production of cholesterol-like molecules impacts Escherichia coli robustness, production capacity, and vesicle trafficking. Santoscoy MC & Jarboe LR. Metabolic Engineering. Link
*Oleic acid based experimental evolution of Bacillus megaterium yielding an enhanced P450 BM3 variant. Vincent T, Gaillet B & Garnier A. BMC Biotechnology. Link
*Triterpenoid production with a minimally engineered Saccharomyces cerevisiae chassis. Ebert BE…Blank LM. bioRxiv (preprint). Link
*Integrated rational and evolutionary engineering of genome-reduced Pseudomonas putida strains empowers synthetic formate assimilation. Turlin J…Nikel PI. bioRxiv (preprint). Link
*Paving the way for synthetic C1- metabolism in Pseudomonas putida through the reductive glycine pathway. Bruinsma L…dos Santos VAPM. bioRxiv (preprint). Link
Tunable and Modular miRNA Classifier through Indirect Associative Toehold Strand Displacement. Chen RP & Chen W. ACS Synthetic Biology. Link
Computational Tools & Models
*Interrogating the effect of enzyme kinetics on metabolism using differentiable constraint-based models. Wilken SE…Ebenhöh O. bioRxiv (preprint). Link
*acCRISPR: An activity-correction method for improving the accuracy of CRISPR screens. Ramesh A…Wheeldon I. bioRxiv (preprint). Link
*Efficient algorithms for designing maximally sized orthogonal DNA sequence libraries. Gowri G, Sheng K & Yin P. bioRxiv (preprint). Link
CRISPR & Gene Editing
*Base editing in human cells with monomeric DddA-TALE fusion deaminases, Geun Mok Y…Kim J-S. Nature Communications. Link
*Enhancement of CRISPR/Cas12a trans-cleavage activity using hairpin DNA reporters. Rossetti M…Porchetta A. Nucleic Acids Research. Link
†Tutorial: design and execution of CRISPR in vivo screens. Braun CJ…Rad R. Nature Protocols. Link
Medicine & Diagnostics
↑↑↑*Cell-free production of personalized therapeutic phages targeting multidrug-resistant bacteria. Emslander Q…Westmeyer GG. Cell Chemical Biology. Link
An inactivated multivalent influenza A virus vaccine is broadly protective in mice and ferrets. Park J…Taubenberger JK. Science Translational Medicine. Link
*Leveraging gene therapy to achieve long-term continuous or controllable expression of biotherapeutics. Cripe TP…Wang P-Y. Science Advances. Link
*A paper-based assay for the colorimetric detection of SARS-CoV-2 variants at single-nucleotide resolution. Zhang T…Li J. Nature Biomedical Engineering. Link
*Targeting oxidized phospholipids by AAV-based gene therapy in mice with established hepatic steatosis prevents progression to fibrosis. Upchurch CM…Leitinger N. Science Advances. Link
*Programmable CRISPR-Cas9 microneedle patch for long-term capture and real-time monitoring of universal cell-free DNA. Yang B, Kong J & Fang X. Nature Communications. Link
*Massively targeted evaluation of therapeutic CRISPR off-targets in cells. Pan X…Luo Y. Nature Communications. Link
*Humanized mice for investigating sustained Plasmodium vivax blood-stage infections and transmission. Luiza-Batista C…Garcia S. Nature Communications. Link
*Engineered ACE2-Fc counters murine lethal SARS-CoV-2 infection through direct neutralization and Fc-effector activities. Chen Y…Pazgier M. Science Advances. Link
*A bioengineered probiotic for the oral delivery of a peptide Kv1.3 channel blocker to treat rheumatoid arthritis. Wang Y…Beeton C. bioRxiv (preprint). Link
Inactivation of a wheat protein kinase gene confers broad-spectrum resistance to rust fungi. Wang N…Wang X. Cell. Link
Protein & Molecular Engineering
↑↑↑Three-dimensional structure-guided evolution of a ribosome with tethered subunits. Soon Kim D…Jewett MC. Nature Chemical Biology. Link
A chemogenetic platform for controlling plasma membrane signaling and synthetic signal oscillation. Suzuki S…Tsukiji S. Cell Chemical Biology. Link
*Structure-based rational design of an enhanced fluorogen-activating protein for fluorogens based on GFP chromophore. Goncharuk MV…Baranov MS. Communications Biology. Link
Tools & Technology
*Magnetically steerable bacterial microrobots moving in 3D biological matrices for stimuli-responsive cargo delivery. Akolpoglu MB…Sitti M. Science Advances. Link
*Precision digital mapping of endogenous and induced genomic DNA breaks by INDUCE-seq. Dobbs FM…Reed SH. Nature Communications. Link
*Building programmable multicompartment artificial cells incorporating remotely activated protein channels using microfluidics and acoustic levitation. Li J…Barrow DA. Nature Communications. Link
Increasing the throughput of sensitive proteomics by plexDIA. Derks J…Slavov N. Nature Biotechnology. Link
De novo designed peptides for cellular delivery and subcellular localisation. Rhys GG…Woolfson DN. Nature Chemical Biology. Link
Real age prediction from the transcriptome with RAPToR. Bulteau R & Francesconi M. Nature Methods. Link
*Swapped genetic code blocks viral infections and gene transfer. Nyerges A…Church GM. bioRxiv (preprint). Link
*Generation of genome-edited dogs by somatic cell nuclear transfer. Kim D-E…Kim M-K. BMC Biotechnology. Link