Week 7 HW: Genetic Circuits Part II
Assignment Part 1: Intracellular Artificial Neural Networks (IANNs)
1. What advantages do IANNs have over traditional genetic circuits, whose input/output behaviors are Boolean functions? Incoherent Analog Neural Networks (IANNs) represent a significant conceptual leap beyond traditional Boolean genetic circuits. Where Boolean circuits reduce biological signals to binary on/off states, IANNs exploit the full continuous, graded nature of molecular concentrations and reaction kinetics, enabling a far richer computational repertoire within living cells. Here is a structured academic overview of their key advantages.
Continuous Signal Processing Over Binary Abstraction Traditional genetic circuits model gene regulation as Boolean functions, forcing inherently graded biochemical signals into discrete 0/1 states. This abstraction discards enormous amounts of biologically encoded information. IANNs, by contrast, operate directly on analog, continuous valued signals, allowing cells to perform proportional, dose responsive computation that mirrors how natural regulatory networks actually function.
Analog computation in synthetic biology offers at least seven distinct benefits over digital approaches, including noise tolerance through stochastic averaging, energy efficiency, and the capacity to represent real valued quantities without the overhead of binary encoding. The continuous regime allows a single network motif to encode functions that would require exponentially many Boolean gates to approximate.
The use of incoherent feedforward loop (IFF) motifs is central to this advantage. IFF architectures generate non monotone, adaptive, and biphasic responses to time varying inputs, behaviors that are structurally impossible in purely Boolean gate designs. These dynamic phenotypes, including overshoots, fold change detection, and transient adaptation, are not artifacts but functional computational primitives that constrain and enrich the network’s input output repertoire.
Expressive Power and Functional Richness Boolean genetic circuits are fundamentally limited to logic operations: AND, OR, NOT, and their combinations. IANNs, drawing on the mathematics of analog neural computation, can approximate arbitrary continuous functions over their input domain, a property rooted in universal approximation theory. This means a single IANN topology can, in principle, implement regression, classification, and dynamic filtering simultaneously.
Synthetic mixed signal computation in living cells demonstrates this directly: by combining analog signal processing with digital switching elements, cells can perform computations such as fuzzy logic inference and neural network style pattern recognition that no purely Boolean circuit can replicate without combinatorial explosion in circuit size.
From a unifying analog circuits perspective, synthetic biology networks including IFF motifs are formally equivalent to analog electronic circuits performing operations such as differentiation, integration, and gain control. This equivalence reveals that Boolean abstraction is not merely a simplification but a genuine loss of computational capacity.
Scalability and Modularity in Multicellular Systems Boolean genetic circuits face severe scalability constraints. As circuit complexity grows, orthogonal parts become scarce, signal degradation accumulates across logic stages, and metabolic burden on the host cell increases nonlinearly. IANNs distribute computation across graded signal levels and can leverage intercellular communication to implement network layers without requiring each cell to host an entire logic chain.
Distributed biological computation frameworks show that multicellular networks, where each engineered strain acts as an analog processing node rather than a discrete logic chip, overcome the modularity bottleneck inherent to monoclonal Boolean designs. The analog regime allows graceful degradation rather than catastrophic failure when individual components are noisy or variable.
Synthetic biological neural networks built on these principles further demonstrate that continuous weight like parameters encoded in promoter strengths and protein binding affinities allow learning analogous behaviors, such as adaptation to environmental statistics, that are categorically outside the scope of fixed Boolean truth tables.
Noise Utilization and Robustness Counterintuitively, IANNs can exploit biological stochasticity rather than being degraded by it. In analog computation, noise can be harnessed through stochastic resonance and averaging effects to improve signal detection at low input amplitudes, a mechanism with no counterpart in Boolean circuits where noise simply causes bit flip errors.
Sarpeshkar’s foundational analysis of analog synthetic biology formalizes this point: the equivalence between probabilistic analog computation and continuous analog computation means that molecular noise is not merely a nuisance to be filtered but a computational resource that can be co opted for robust, energy efficient information processing. 2. Describe a useful application for an IANN; include a detailed description of input/output behavior, as well as any limitations an IANN might face to achieve your goal. One of the most compelling and clinically meaningful applications for an IANN is intracellular multi-biomarker cancer classification, where an engineered cell continuously reads several molecular signals simultaneously and produces a graded therapeutic output proportional to the likelihood and severity of a malignant state. This application exploits precisely the computational properties that Boolean circuits cannot provide, and it also surfaces the sharpest limitations IANNs currently face. Here is a full structured treatment.
The Application: Intracellular Analog Cancer Classifier The core idea is to engineer a living cell, plausibly a T cell or a synthetic probiotic, to act as a roving analog classifier. Rather than toggling a single kill switch when one biomarker crosses a hard threshold, the cell continuously integrates the concentrations of multiple tumor associated molecular inputs and produces a continuously graded output, such as the expression level of a therapeutic effector protein, that scales with the computed probability of malignancy. This is precisely the paradigm described in synthetic mixed signal computation, where analog to digital converters built from comparator gene circuits can digitize graded inputs at multiple thresholds, but the richer IANN regime goes further by keeping the output in the analog domain throughout, allowing proportional dosing rather than binary actuation.
Molecular computation frameworks confirm that both Boolean logic and analog neural network architectures have been explored for molecular classification, but the analog neural network route is uniquely suited to problems where the decision boundary in biomarker space is nonlinear and continuous, as is the case in cancer, where no single biomarker cleanly separates malignant from healthy tissue.
Input Behavior The inputs to the IANN are the intracellular or extracellular concentrations of several co-expressed tumor biomarkers, for example HER2, MUC1, and a hypoxia responsive transcription factor such as HIF-1α. Each of these signals is a continuous, real valued quantity that fluctuates over time and varies across cell populations. The IANN does not binarize these signals at the input layer. Instead, each input drives a dedicated promoter whose transcriptional output is a graded, monotone increasing function of the corresponding ligand concentration, effectively implementing a continuous input encoding layer.
The incoherent feedforward motifs within the network then operate on these graded inputs simultaneously. Because IFF motifs generate non-monotone, adaptive responses, including overshoots and fold change detection, the network is sensitive not only to the absolute concentration of each biomarker but also to the rate of change and relative ratio between inputs. This is a critical functional advantage: a tumor cell may express HER2 at a moderate absolute level but show a sharply rising trajectory, and the IFF architecture captures that dynamic signature in a way that a static Boolean threshold cannot. Sontag’s input output theory of dynamic phenotypes formalizes this precisely, showing that IFF motifs encode fold change detection and biphasic adaptation as structural, parameter independent properties of the network topology.
Output Behavior The output of the IANN is the steady state or transient expression level of a therapeutic protein, for instance a cytokine, a pro-apoptotic factor, or a checkpoint inhibitor. This output is a continuous, weighted, nonlinear function of all inputs processed through the network’s layered IFF motifs. Critically, the output is graded: low combined biomarker scores produce low effector expression, intermediate scores produce intermediate expression, and high scores produce maximal therapeutic output. This proportionality is what makes the IANN clinically superior to a Boolean kill switch, which either fires fully or not at all, creating risks of both under-treatment and toxic over-activation.
Enzymatic neural network experiments published in Nature directly validate this output regime. In those systems, analog neurons whose computation varies continuously with input concentration were shown to perform nonlinear decision making, with the incoherent feedforward architecture specifically identified as the mechanism generating the non-monotone output surface needed for classification in overlapping signal spaces. The output layer of such a network can be tuned by adjusting enzyme concentrations, which act as synaptic weights, to shift the decision boundary without redesigning the circuit topology.
The mixed signal framework further shows that when a continuous output must eventually trigger a discrete cellular behavior, an analog to digital interface layer can be appended downstream, allowing the IANN to drive a sharp effector response only when the analog score crosses a final confidence threshold, combining the best of both computational regimes.
Limitations Despite this promise, IANNs face several serious and well documented limitations in achieving this application.
Parameter sensitivity and tunability. The weights of a biological IANN are encoded in promoter strengths, ribosome binding site efficiencies, and protein degradation rates. Unlike silicon synaptic weights, these parameters cannot be set with numerical precision. Small variations in plasmid copy number or host cell metabolism shift the analog decision boundary in ways that are difficult to predict or correct. Machine learning and deep learning surveys of synthetic biology explicitly identify this as a central challenge: the mapping from designed sequence to quantitative analog behavior remains unreliable enough to require extensive experimental iteration for each new input combination.
Context dependence and metabolic burden. An IANN with multiple input channels and several IFF layers places substantial transcriptional and translational load on the host cell. This metabolic burden alters the very kinetic parameters the network depends on, creating a feedback between circuit operation and circuit performance that degrades analog fidelity. Deep learning frameworks applied to synthetic biology highlight continuous chemostat control as a partial mitigation strategy, but in a therapeutic cell operating in vivo, no such external control is available.
Noise and cell to cell variability. While population level analog averaging can be exploited, individual cell to cell variability in protein expression means that the IANN output distribution across a clonal population is broad. For a therapeutic application, this translates into a distribution of effector doses across individual cells, some of which may fall below therapeutic threshold and others above toxic threshold simultaneously. The stochastic resonance benefits of analog computation help at the population mean but do not eliminate the tail risks at the single cell level.
Lack of in vivo training. Unlike a silicon neural network, a biological IANN cannot be retrained after deployment. The weights are fixed at the time of cell engineering. If the tumor evolves its biomarker expression profile, as cancers routinely do under selective pressure, the classifier’s decision boundary becomes misaligned with the new signal space and therapeutic efficacy degrades without any mechanism for correction.
3. Below is a diagram depicting an intracellular single-layer perceptron where the X1 input is DNA encoding for the Csy4 endoribonuclease and the X2 input is DNA encoding for a fluorescent protein output whose mRNA is regulated by Csy4. Tx: transcription; Tl: translation.
Draw a diagram for an intracellular multilayer perceptron where layer 1 outputs an endoribonuclease that regulates a fluorescent protein output in layer 2.
The diagram shows a two-input biological perceptron where X1 (Csy4 DNA) and X2 (fluorescent protein DNA) are each transcribed by their own Tx nodes. Inside the computational soma (dashed circle), Csy4 protein translated from X1 mRNA acts as a negative synaptic weight by cleaving the X2 mRNA via its endoribonuclease activity, while X2 contributes a positive excitatory drive. The net surviving X2 mRNA exits the node, gets translated by the downstream Tl node, and produces a continuously graded fluorescent output Y whose intensity is a smooth, analog function of both input concentrations simultaneously. High X1 suppresses Y; high X2 elevates Y; and intermediate combinations yield proportional intermediate outputs, making this a real-valued weighted subtraction implemented entirely in molecular kinetics rather than Boolean switching.
Assignment Part 2: Fungal Materials
1. What are some examples of existing fungal materials and what are they used for? What are their advantages and disadvantages over traditional counterparts? Fungal mycelium materials, grown by binding agricultural waste substrates with the thread-like root networks of fungi, have emerged across several industries as credible alternatives to conventional synthetic and petrochemical products. The most prominent examples include mycelium-based composites (MBCs) used as packaging foam replacements (most famously Ecovative Design’s Mushroom Packaging), thermal and acoustic insulation panels substituting expanded polystyrene, leather-like textiles such as Bolt Threads’ Mylo replacing animal hide, and structural construction blocks being explored as substitutes for low-load-bearing concrete and timber panels. In food systems, filamentous fungal mycelium is also processed into protein-rich meat analogues, replacing conventional animal protein sources.
On the advantages side, MBCs are biodegradable, carbon-neutral in production, grown at ambient temperature without energy-intensive manufacturing, and use agricultural byproducts like rice husk, hemp, and bamboo sawdust as feedstock, dramatically reducing waste. They are lightweight, naturally fire-resistant to a degree, and non-toxic throughout their lifecycle, giving them a decisive environmental edge over polystyrene foam, synthetic leather, and fiberglass insulation, all of which carry substantial ecological costs in production and disposal.
The disadvantages, however, are equally real. MBCs exhibit high moisture sensitivity, absorbing water and losing mechanical integrity in humid environments, which is their most cited structural weakness relative to conventional counterparts. Their compressive and tensile strength remains below that of plastics and concrete for most species and substrate combinations, limiting them to non-load-bearing applications. Production is slower and less scalable than industrial polymer manufacturing, and batch-to-batch variability in mycelial growth introduces inconsistency that synthetic materials do not have. Thermal insulation performance, while competitive, still falls short of high-performance synthetic foams in extreme temperature ranges.
2. What might you want to genetically engineer fungi to do and why? What are the advantages of doing synthetic biology in fungi as opposed to bacteria? Genetically engineering fungi opens a remarkably broad application space. The most compelling targets include engineering mycelium to overproduce specific structural proteins or chitin-binding polymers to enhance the mechanical and moisture resistance of mycelium-based materials, programming fungi to biosynthesize high-value pharmaceuticals and natural products such as penicillin derivatives, statins, and novel antimicrobials through genome-mined biosynthetic gene clusters, optimizing filamentous fungi like Aspergillus and Trichoderma for precision fermentation of alternative proteins and food ingredients, and engineering biopesticide fungi such as Beauveria bassiana with enhanced virulence against specific insect pests while reducing off-target effects. Beyond materials and food, fungi are also being engineered for next-generation biofuel production, leveraging their native lignocellulose-degrading machinery to convert agricultural waste directly into ethanol and other energy carriers.
The advantages of doing synthetic biology in fungi rather than bacteria are substantial and multifaceted. Fungi are eukaryotes, meaning they possess the post-translational modification machinery, including glycosylation, disulfide bond formation, and proper protein folding via the endoplasmic reticulum, that bacteria entirely lack. This makes fungi far superior hosts for producing complex eukaryotic proteins, including many therapeutics and food proteins, that are misfolded or non-functional when expressed in E. coli. Fungi also tolerate and actively process large, intron-containing genes and complex biosynthetic gene clusters that are too large or toxic to stably maintain in bacterial hosts. Their filamentous growth morphology allows them to physically penetrate and colonize solid substrates, enabling applications in materials, soil remediation, and biopesticides that are structurally impossible for planktonic bacteria. CRISPR-based tools have now been adapted specifically for fungal genomes, enabling precise multiplex editing across both yeast and filamentous species with efficiency approaching that long available in bacteria.
The tradeoffs are real, however. Fungi grow more slowly than bacteria, their genetic toolkits are less mature and less standardized, transformation efficiency is lower in many filamentous species, and regulatory frameworks for releasing engineered fungi into open environments are considerably more complex than for bacterial systems. Nevertheless, for applications demanding eukaryotic biology, structural complexity, or substrate-invasive growth, fungi represent a uniquely powerful chassis that bacteria simply cannot replicate.
Assignment Part 3: First DNA Twist Order
Review the Individual Final Project documentation guidelines. Submit this Google Form with your draft Aim 1, final project summary, HTGAA industry council selections, and shared folder for DNA designs. DUE MARCH 20 FOR MIT/HARVARD/WELLESLEY STUDENTS Review Part 3: DNA Design Challenge of the week 2 homework. Design at least 1 insert sequence and place it into the Benchling/Kernel/Other folder you shared in the Google Form above. Document the backbone vector it will be synthesized in on your website.