Video Presentations

Recurrent neural chemical reaction networks that approximate arbitrary dynamics

Many important phenomena in biochemistry and biology exploit dynamical features such as multi-stability, oscillations, and chaos. Construction of novel chemical systems with such rich dynamics is a challenging problem central to the fields of synthetic biology and molecular nanotechnology. In this work, we address this problem by putting forward a molecular version of a recurrent artificial neural network, which we call the recurrent neural chemical reaction network (RNCRN). The RNCRN uses a modular architecture - a network of chemical neurons - to approximate arbitrary dynamics.

We prove that with sufficiently many chemical neurons and suitably fast reactions, the RNCRN can be systematically trained to achieve any well-behaved dynamics. RNCRNs with relatively small number of chemical neurons and a moderate range of reaction rates are then trained on ordinary differential equations (ODEs) to display a variety of biologically-important dynamical features including oscillations, robustness, and bifurcations. We then introduce an algorithm for producing designer dynamics without an underlying ODE model, simply by specifying desired dynamical features. We produce chemical reaction networks with irregular dynamic behaviours including a heart-shaped attractor, disjoint attractors, and a toroidal attractor.

We then train chemical systems to toggle between two data-defined target dynamical behaviours. We show that the switching behaviour of these hybrid systems can be trained to depend on non-linear conditions of upstream chemical parameters leading to abstract chemical systems reminiscent of the regulatory framework found in life. Finally, we argue that small RNCRNs are experimentally implementable with DNA-strand-displacement technologies and discuss implementation approaches more broadly in synthetic biology.

The physical limits of molecular templating

The systems of the biological cell achieve awesome and inspirational feats of molecular assembly. A paradigmatic example is the selective assembly of essentially arbitrary functional RNA and protein molecules from only a small number of building blocks. This process is made possible by the copying of information from a DNA template sequence into the sequence of daughter polymers (RNA and then proteins).

Templating is essential in creating biochemical complexity in living systems, but it is almost entirely ignored in synthetic contexts. We argue that this oversight largely results from the fact that templating is necessarily an extraordinarily far from equilibrium process, making it hard to engineer. We also ask: given that accurate templating results in a far from equilibrium system, how much chemical work must be put in to maintain such a state in an arbitrary system in which products are continuously produced and degraded? We find that the accuracy of the product ensemble is upper bounded, by a function of ΔG, the difference between the maximal and minimal free-energy changes along pathways to product assembly. Remarkably, however, although ΔG constrains the information propagated to the product distribution, the systems that saturate the bound do not look like their biological counterparts, instead operating in a pseudo-equilibrium fashion, with production and degradation for each product sequence largely occurring via the same pathway in forward and reverse directions, rather than through the free-energy consuming cycles observed in biology. Indeed, the larger the cyclic flux observed in the system, the worse the precision. This surprising result raises the question of why biology operates in the limit of large cyclic flux, and also suggests a possible low-energy paradigm for molecular computation.

Implementing Non-Equilibrium Networks with Active Circuits of Duplex Catalysts

DNA strand displacement (DSD) reactions have been used to construct chemical reaction networks in which species act catalytically at the level of the overall stoichiometry of reactions. These effective catalytic reactions are typically realised through one or more of the following: many-stranded gate complexes to coordinate the catalysis, indirect interaction between the catalyst and its substrate, and the recovery of a distinct "catalyst" strand from the one that triggered the reaction. These facts make emulation of the out-of-equilibrium catalytic circuitry of living cells more difficult. Here, we propose a new framework for constructing catalytic DSD networks: Active Circuits of Duplex Catalysts (ACDC). ACDC components are all double-stranded complexes, with reactions occurring through 4-way strand exchange. Catalysts directly bind to their substrates, and the "identity" strand of the catalyst recovered at the end of a reaction is the same molecule as the one that initiated it. We analyse the capability of the framework to implement catalytic circuits analogous to phosphorylation networks in living cells. We also propose two methods of systematically introducing mismatches within DNA strands to avoid leak reactions and introduce driving through net base pair formation. We then combine these results into a compiler to automate the process of designing DNA strands that realise any catalytic network allowed by our framework.

https://drops.dagstuhl.de/opus/volltexte/2020/12960/pdf/LIPIcs-DNA-2020-7.pdf

 Avoiding equilibrium in a minimal molecular information-processing system

Non-Equilibrium Thermodynamics of Catalytic Information Processing

Tom's talk at the Banff International Research Station workshop "Mathematical Models in Biology: from Information Theory to Thermodynamics" (2020).

Catalytic motifs are ubiquitous in cellular information-processing systems, from kinase signalling networks to the central dogma of molecular biology. This ubiquity results from the ability of catalysts to channel chemical free energy into far-from-equilibrium information-bearing states, allowing them to perform non-trivial computational operations. This power, however, comes at a price. At a fundamental level, the need to create non-equilibrium outputs sets thermodynamic constraints on these systems. At a practical level, catalysts must carefully balance kinetic and thermodynamic factors to ensure that the desired non-equilibrium output is actually reached. The complexity of this task explains the comparatively slow progress made with engineering synthetic non-equilibrium information-processing systems, as opposed to synthetic systems that form complex equilibrium assemblies. I will present our latest work - both theoretical and experimental - aimed at overcoming this challenge to engineer non-equilibrium catalytic systems for information processing.

Papers from which this talk is drawn:

https://link.springer.com/article/10.1007/s11047-017-9646-x

https://arxiv.org/abs/1905.00555

https://www.biorxiv.org/content/10.1101/2020.05.22.108571v2

Handhold-mediated strand displacement for non-equilibrium templating

Toehold-mediated strand displacement (TMSD) is a nucleic acid-based reaction wherein an invader strand (I) replaces an incumbent strand (N) in a duplex with a target strand (T). TMSD is driven by toeholds, overhanging single-stranded domains in T recognised by I. Although TMSD is powerful, it cannot implement templating, the mechanism by which biological systems generate far-from equilibrium assemblies like RNA or proteins. Therefore, we introduce handhold-mediated strand displacement (HMSD). Handholds are toehold analogues located in N and capable of implementing templating. We demonstrate that handholds can accelerate the rate of invader-target (IT) binding by more than 4 orders of magnitude. Furthermore, handholds of moderate length accelerate reactions whilst allowing detachment of the product IT from N. We are thus able to experimentally demonstrate the use of HMSD-based templating to produce highly-specific far-from-equilibrium DNA duplexes.

https://www.biorxiv.org/content/10.1101/2020.05.22.108571v1

A broad intro to Javi's work for Imperial's three-minute-thesis competition.