Evolution and evolvability · Article

Auto- and heteroassociative memory in gene regulatory networks

Associative memory is the capacity of a system to link one stimulus or pattern to another and to recall it later. Watson and colleagues compare gene networks to the associative memory of a Hopfield network. Several authors argue that gene regulatory networks can do the same, with regulatory weights standing in for synaptic strengths. On this view development can store and recall phenotypes selected in the past.⁵⁷

Earliest held: 1890, James, W.
Most discussed in: Comparative Physiology of the Brain and…, 1900
In the library: 134 passages in 28 works
Rewritten: 2026-10-26

Behavioural criterion

Early uses of the term appear in comparative and animal psychology. Loeb (1900) proposed that an animal which can be trained to react to certain stimuli must possess associative memory, and he offered this as a criterion for comparative psychology.¹ Jennings (1906) described it objectively as a response first given to one stimulus that comes to be transferred to another.² Jennings tied it to the same law of resolving physiological states that underlies ordinary memory. Loeb (1918) found no adequate proof of it in worms, starfish, medusae or infusorians, and linked it to definite parts of the brain.³

Regulatory weights

Gene circuit models give networks of genes a form close to that of neural networks. Crombach and colleagues (2012) describe a gap gene model in which each entry of an interconnectivity matrix is a regulatory weight. A weight can mean activation, repression or no interaction, depending on whether it is positive, negative or near zero.⁴ Szilágyi and colleagues (2020) state that control genes form network-like architectures that are mathematically equivalent to artificial neural networks, citing Vohradsky.⁸

Developmental memory

Watson and colleagues (2016) argued that regulatory interactions evolve under natural selection in the same manner as correlation learning changes synaptic connections. They concluded that gene networks can hold a developmental memory of several past phenotypes, like a Hopfield network.⁵ They added that such networks can produce novel phenotypes that share structural regularities with past ones. De la Fuente and colleagues (2019) made a related claim for metabolism. Their earlier modelling suggested that modular metabolic networks follow Hopfield-like attractor dynamics, and their amoeba experiments were inspired by that prediction.⁶

Auto- and heteroassociation

Szilágyi and colleagues (2020) extended gene network theory to cover both auto- and heteroassociativity.⁷ They noted that the theory of autoassociative networks cannot yet be carried over directly to development. Standard models use positive and negative state variables, while developmental systems work with nonnegative concentrations.⁸ Biswas, Clawson and Levin (2022) raised a separate question. Earlier work had not shown whether gene regulatory networks or protein pathways could be trained without erasing prior experience, and so whether one network could keep more than one memory.⁹

SourcesEach quotation was checked word for word against the passage it opens.

if an animal can learn, that is, if it can be trained to react in a desired way upon certain stimuli (signs), it must possess associative memory.Loeb, J., 1900 · Comparative Physiology of the Brain and Comparative Psychology · open at passage 320
This is characterized objectively by the fact that the response at first given to one stimulus comes, after a time, to be transferred to another one.Jennings, H. S., 1906 · Behavior of the Lower Organisms · open at passage 938
to the knowledge of the writer no adequate proof for its existence has ever been given for worms, starfish, sea urchins, actinians, medusae, hydroids, or infusorians.Loeb, J., 1918 · Forced Movements, Tropisms, and Animal Conduct · open at passage 293
each regulatory weight in this matrix can represent activation, repression, or no interaction depending on whether it is positive, negative or (close to) zeroCrombach A, Wotton KR, Cicin-Sain D, Ashyraliyev M, Jaeger J, 2012 · Efficient reverse-engineering of a developmental gene regulatory network · open at passage 5
gene-regulation networks can exhibit a “developmental memory” able to store and recall multiple phenotypes that have been selected for in the past, exactly like the associative memory of a Hopfield network.Watson RA, Mills R, Buckley CL, Kouvaris K, Jackson A…, 2016 · Evolutionary Connectionism: Algorithmic Principles Underlying the Evolution of… · open at passage 72
self-organized enzymatic activities in modular metabolic networks seem to be governed by Hopfield-like attractor dynamics similar to what happens in neural networksDe la Fuente IM, Bringas C, Malaina I, Fedetz M…, 2019 · Evidence of conditioned behavior in amoebae · open at passage 44
In the present investigation we extend the theory of gene regulatory networks to involve both auto- and heteroassociativity.Szilágyi A, Szabó P, Santos M, Szathmáry E, 2020 · Phenotypes to remember: Evolutionary developmental memory capacity and… · open at passage 5
in ontogenetic systems the relevant space is that of nonnegative real numbers, corresponding to concentrations of different moleculesSzilágyi A, Szabó P, Santos M, Szathmáry E, 2020 · Phenotypes to remember: Evolutionary developmental memory capacity and… · open at passage 4
However, in prior work it was not known whether GRNs or protein pathways could be trained in a way that did not wipe out the effects of prior experience.Biswas S, Clawson W, Levin M, 2022 · Learning in Transcriptional Network Models: Computational Discovery of… · open at passage 22

Linked ideas

Gene regulatory network models of associative learning and bistable memoryrelated to

Treats the same idea of gene networks acting as associative memories; read together for the modelling detail.

Developmental memory capacity, robustness and attractor basins in gene networksrelated to

Covers the evolutionary developmental memory work that the auto- and heteroassociative extension builds on.

Homeostatic learning rules for gap junction couplingparallels

Applies learning rules to a different substrate, gap junction networks, rather than to regulatory weights in gene networks.

Gene regulatory networks as Boolean networksrelated to

Another formal treatment of gene networks, using Boolean states rather than the nonnegative concentrations discussed here.

Gap gene network and positional informationprecursor of / follows

Gene circuit models of the gap gene network supply the regulatory weight matrices used in this analogy.

Associative learning in plantsparallels

Asks the behavioural question of associative learning in organisms without nervous systems.

Where it is discussedPassages matching autoassociativity, heteroassociative, associative memory, Hopfield network

1900	Comparative Physiology of the Brain and Comparative Psychology · Loeb, J.	51
2016	Evolutionary Connectionism: Algorithmic Principles Underlying the Evolution of… · Watson RA, Mills R, Buckley CL…	7
2020	Phenotypes to remember: Evolutionary developmental memory capacity and… · Szilágyi A, Szabó P, Santos M…	6
1912	The Mechanistic Conception of Life · Loeb, J.	6
2015	What can ecosystems learn? Expanding evolutionary ecology with learning theory · Power DA, Watson RA, Szathmáry E, Mills…	5
1918	Forced Movements, Tropisms, and Animal Conduct · Loeb, J.	5
2012	Efficient reverse-engineering of a developmental gene regulatory network · Crombach A, Wotton KR, Cicin-Sain D…	5
2008	Retention of memory through metamorphosis: can a moth remember what it learned… · Blackiston DJ, Silva Casey E, Weiss MR	5
1908	The Animal Mind: A Textbook of Comparative Psychology · Washburn, M. F.	4
2013	Stochasticity, bistability and the wisdom of crowds: a model for associative… · Sorek M, Balaban NQ, Loewenstein Y	4