Bacterial Reservoir Computer

In contrast to traditional bottom-up approaches that build biological devices for computation within organisms [1-3], this PhD project proposes a novel top-down strategy. The goal is to leverage bacterial strains in a reservoir computing (RC) framework to solve complex computational tasks.

Engineering bottom-up biological devices is challenging. These devices place a significant metabolic burden on host cells, are difficult to fine-tune, and are prone to noise [4, 5]. The design of such devices often draws inspiration from biological information-processing systems, akin to logic gates, switches, and perceptron, already found in nature [6, 7]. This raises an intriguing question: instead of constructing devices from the ground up, could natural microorganisms themselves be harnessed for complex computational tasks?

By developing a reservoir computing approach with bacterial strains, this PhD research is expected to highlight the potential of a top-down approach in synthetic biology for biocomputing, with implications for solving complex tasks typically handled by digital systems. The project also aims to set the groundwork for further applications in medical diagnostics and propose avenues for integrating bacterial reservoir computing with emerging technologies, such as neuromorphic computing and engineered living materials.

Objective and context

Reservoir computing (RC) is a branch of artificial intelligence exploring the computational capabilities of physical, chemical, and biological systems [8]. Initially developed as an alternative to classical artificial neural networks, particularly recurrent neural networks (RNNs), RC offers a more efficient training process. There are two kinds of RC systems: conventional RC and physical RC.

In a typical conventional RC framework, input data is provided to a reservoir, the states of the reservoir nodes are read, and this information is fed into a post-readout feedforward network—usually a simple linear classifier or regressor—that interprets the reservoir states to produce the final output. Unlike RNNs with trainable weights, conventional RC uses a fixed-weight RNN reservoir, making it faster to train and suitable for applications such as time series prediction and forecasting, dynamic system control, and Internet of Things (IoT) applications where quick, efficient processing is crucial [9].

In physical RC, the reservoir is replaced by a physical object (Figure 1). The system operates by first feeding input data into the physical reservoir, which processes the inputs and transforms them into a high-dimensional, dynamic state. These transformed states are then passed to a post-readout layer, similar to conventional RC. Physical RC has many applications, including the seminal liquid state machine for pattern recognition in a bucket of water [10], chemical RC developments for classification tasks and forecasting solutions of ODE systems [11], and biological RC systems using rat cortical neurons cultured on micropatterned substrates to solve classification tasks [12].

This PhD project aims to investigate the feasibility of using bacterial strains within an RC framework, assessing their potential as reservoirs for computation. Like other machine learning methods, RC relies on training data with features and labels, seeking to predict labels from features. In a bacterial reservoir approach, problem features are represented as nutrients provided to the bacteria, and bacterial responses are measured through phenotypic observations. These measurements are then processed by classical machine learning regressor or classifier to produce solutions to computational tasks.

Methods and Work Plan

Practically, the project will begin (Year 1) by using E. coli as a test strain, developing a hybrid model  trained on media supplemented with different metabolites and acquiring growth curves. Various neural mechanistic models will be benchmarked among these Physics Informed Neural Networks (PINNs [13]), neural – flux balance analysis (FBA) [14] and neural dynamic-FBA. The computational capabilities of this E. coli reservoir will then be benchmarked against classical machine learning techniques, such as Multiple Linear Regression (MLR), Support Vector Machine (SVM), and Multilayer Perceptron (MLP), across various regression and classification tasks. As shown in Faulon et al. [15], preliminary results obtained on a set of E. coli growth rates acquired for different nutrients (i.e., different sets of sugars, amino acids, nucleotides) indicate such approach can be used to classify linear and non-linear patterns.

As a practical application, bacterial reservoirs will also be tested (Year 2) for classification of clinical samples, using wild-type or mutant bacterial strains. Preliminary data suggest that mutant E. coli reservoir can accurately classify COVID-19 samples, differentiating between moderate and severe cases based on strain responses.

In Year 3, reservoirs based on genome-scale metabolic models (GEMs) developed in previous years for several bacterial species or mutant (gene KO) species will be explored, growth data will be experimentally collected for the most promising candidates. The collected species and their growth data will then serve to solve classical regression and classification tasks along with testing on clinical samples. Here instead of using a single species, the multi-species RC system will be composed of a set of wild-type or mutant species acting as reservoirs.

For all testing on clinical samples the PhD student will have access to prostate cancer cohort and Covid-19 cohort provided by the University Hospitals of Montpelier and Grenoble.  Additional samples for other diseases may be acquired during the course of the project. The possibility of monitoring environmental pollutants (for instance in water) will also be investigated using the multi-species RC framework.

Further information on the subject: https://jfaulon.com/phd-subject-bacterial-reservoir-computer/

References

1.     Purnick, P. E. M. & Weiss, R. The second wave of synthetic biology: from modules to systems. Nat. Rev. Mol. Cell Biol. 10, 410–422 (2009).

2.     Gardner, T. S., Cantor, C. R. & Collins, J. J. Construction of a genetic toggle switch in Escherichia coli. Nature 403, 339–342 (2000).

3.     Elowitz, M. B. & Leibler, S. A synthetic oscillatory network of transcriptional regulators. Nature 403, 335–338 (2000).

4.    Swain, P. S., Elowitz, M. B. & Siggia, E. D. Intrinsic and extrinsic contributions to stochasticity in gene expression. Proc. Natl. Acad. Sci. 99, 12795–12800 (2002).

5.    Borkowski, O., Ceroni, F., Stan, G.-B. & Ellis, T. Overloaded and stressed: whole-cell considerations for bacterial synthetic biology. Curr. Opin. Microbiol. 33, 123–130 (2016).

6.    Hellingwerf, K. J., Postma, P. W., Tommassen, J. & Westerhoff, H. V. Signal transduction in bacteria: phospho-neural network(s) in Escherichia coli ? FEMS Microbiol. Rev. 16, 309–321 (1995).

7.    Scheres, B. & Van Der Putten, W. H. The plant perceptron connects environment to development. Nature 543, 337–345 (2017).

8.    Tanaka, G. et al. Recent advances in physical reservoir computing: A review. Neural Netw. 115, 100–123 (2019).

9.    Chen, H. et al. Emerging memristors and applications in reservoir computing. Front. Phys. 19, 13401 (2024).

10. Fernando, C. & Sojakka, S. Pattern Recognition in a Bucket. in Advances in Artificial Life (eds. Banzhaf, W., Ziegler, J., Christaller, T., Dittrich, P. & Kim, J. T.) vol. 2801 588–597 (Springer Berlin Heidelberg, Berlin, Heidelberg, 2003).

11. Baltussen, M. G., De Jong, T. J., Duez, Q., Robinson, W. E. & Huck, W. T. S. Chemical reservoir computation in a self-organizing reaction network. Nature 631, 549–555 (2024).

12. Sumi, T. et al. Biological neurons act as generalization filters in reservoir computing. Proc. Natl. Acad. Sci. 120, e2217008120 (2023).

13. Yazdani A, et al.  Systems biology informed deep learning for inferring parameters and hidden dynamics. PLoS Comput Biol 16(11), e1007575 (2020).

14. Faure L, et al., A neural-mechanistic hybrid approach improving the predictive power of genome-scale metabolic models. Nat Commun 14, 4669 (2023).

15. Faulon, JL. et al. Reservoir Computing with bacteria. bioRxiv DOI: 10.1101/2024.09.12.612674 (2024).

MICALIS (INRAE & University of Paris Saclay, https://www.micalis.fr) is a research unit of more than 350 researchers developing multidisciplinary approaches and promoting the development of synthetic biology applications for health and biotechnology. Within MICALIS , the recruiting research team (~20 staff spread into wet and dry labs www.jfaulon.com) specializes in developing computer-aided design tools for biotechnology. The PhD student will closely collaborate with molecular biologists and software developers and interact with Research and Data Scientists, Engineers, and other Ph.D. fellows.

M2 research in computational biology , bioinformatics, applied mathematics, or another relevant discipline. Passion for programming and machine learning. Experience with Python.