Learning involves goals and error-assessment mechanisms
At its simplest level, whole organism learning requires two things: (1) a goal (or set point), usually determined in advance, and (2) an error-indicating mechanism that quantifies how close newly changed behaviour approaches that goal. For those who prefer a familiar human example with a short-term goal, learning to ride a bike is a good model. The process of learning requires a continual exchange of information and feedback from the goal to the current behaviour in order to correct current behaviour and direct future behaviour more closely towards achieving the goal.
Wild plants need trial-and-error learning because the environmental circumstances in which signals arrive can be so variable. That is, the starting point can be indeterminate and rote behaviour would be insufficient to ensure successful progress towards the goal. Whereas the eventual fitness goal may always be the same, the life trajectories attempting to achieve that goal must be learnt. Indications of trial-and-error learning can be deduced from the presence of damped or even robust oscillations in behaviour as the organism continually assesses and makes further corrections to behaviour. The reason that plants respond to gravity, for example, is primarily one of nutrition (shoots to light, roots to minerals and water), leading to better growth and eventual reproduction. But roots and shoots may find themselves at any angle to the final desired position and thus must learn progressively how to approach the internally specified optimal angle if conditions allow. However, the final branch angle adopted depends on a congruence of environmental assessments with internally specified information which can be accessed as a default position when conditions are optimal.
There are numerous plant learning examples, and I detail a few to indicate the point. Oscillations and overshoot in the approach of seedling shoots or roots to the vertical after horizontal displacement have been reported, for example, by Johnsson and Israelsson (1968); Heathcote and Aston (1970); Shen-Miller (1973); and Ishikawa et al. (1991). Johnsson (1979) lists a further 23 earlier references that report this behaviour. Bennet-Clerk and Ball (1951) detailed the gravitropic behaviour of many individual rhizomes and report overshoot, undershoot, growth initially in the wrong direction and sustained oscillations. These authors specifically note that averaging tends to eliminate detection of individual behaviour because individuals are rarely in synchrony with each other. Clifford et al. (1982) reported that deliberate bending of Taraxacum shoots causes over-compensatory growth in the other direction upon release, again indicating error correction with a goal (or set point).
Bose (1924) used continuous recording to report that the behaviour of petioles, roots, styles and leaflets of Mimosa to thermal, mechanical and light stimuli often oscillated in their approach to a new state of growth.
When leaves are deprived of water, stomata reduce aperture size, but a tendency to overshoot and oscillations in the new steady state have both been reported (Stalfelt, 1929, quoted in Raschke, 1979). Raschke (1970) detected oscillations of the average stomatal aperture determined by porometry in different regions of maize leaves. Johnsson (1976) concluded that both feedback and feed-forward mechanisms are involved in error correction and optimizing stomatal aperture.
Following mild water stress there is often a period of compensatory growth after rewatering, indicating an error-correction mechanism (Stocker, 1960). Trees can abscind sufficient leaves to adjust numbers to current water supplies. Some trial-and-error mechanism must determine when sufficient have been dropped (Addicott, 1982). Similar mechanisms must be present for all phenotypically plastic processes. Thus, for example, stem thickening in response to wind sway must be able to access the goal of optimal wind sway and a trial-and-error assessment of how far the individual is from that goal.
Resistance to drought or cold can be enhanced by prior treatment to milder conditions of water stress or low temperature (e.g. Kramer, 1980; Kacperska and Kuleza, 1987; Griffiths and McIntyre, 1993). Such well-known behaviour (acclimation) requiring physiological and metabolic changes is analogous to animal learning.
Similarities in avoidance responses by plants and animals
A single stimulus in the marine snail, Aplysia, designed to produce avoidance responses (the goal in this case) may only initiate short-term memory changes lasting a few minutes (Kandel, 2001). The intracellular mechanisms involve the second messengers Ca2+ and cyclic nucleotides and a limited number of protein kinases that phosphorylate ion channels that serve as temporary memory (Greengard, 2001). Repetition of the stimulus or increasing its intensity modifies protein synthesis in neurones and the formation of new dendrites (neural connections). The transduction of these avoidance stimuli involves MAP kinases, control of gene expression by cyclic nucleotide binding elements (CREB), and the ubiquitin pathway to dispose of protein kinase A-regulatory proteins. Increasing the size of the stimulus again greatly enhances further dendrite formation and results in a strengthening and increased effectiveness of dendrites already present in the chosen pathway of communication by adhesion mechanisms that may involve integrins. Additional growth factors are now involved including EF1 (Greengard, 2001), a protein with similar functions in both animals and plants. The new dendrites in this animal represent memory and as they disappear so the memory disappears.
Drought avoidance behaviour by plants is well established. Slight variations in water availability incur equally slight, but temporary, reductions only in cell growth rate, probably involving changes in second messengers, particularly cytosolic Ca2+, [Ca2+]i, and phosphorylation changes in turgor-generating ATPases and associated ion channels (Begg, 1980; Hanson and Trewavas, 1982; Palmgren, 2001). More intense stress signals initiate changes in protein and wall synthesis, cuticle thickness, stomatal conductance and limited morphological reductions of leaf area (Hsaio et al., 1976; Kramer, 1980). Each of these processes seems to have a discrete water potential threshold at which it is initiated. Perhaps progressive reductions in plasma membrane wall adhesion are responsible, initiating transduction mechanisms and modifying plasmodesmatal functioning. The transduction mechanisms include those mentioned above and MAP kinases and other protein kinases modifying transcription factors (Hetherington, 2001; Jonak et al., 2002).
With more severe water stress, the root : shoot ratio increases and, in wild plants, it can vary up to 20-fold (Chapin, 1980). In developing leaves, the internal mesophyll surface area is reduced and stomatal density modified, producing a xeromorphic-type morphology (Stocker, 1960). Increased hairiness, early flowering and a modified vascular system are induced later, indicative of memory of the initial droughting signal (Stocker, 1960; Kramer, 1980).
All of the above responses, whether physiological or morphological, must be initiated and transduced by mechanisms that can assess the current supply of water against a notional optimal supply. The plant learns by trial and error when sufficient changes have taken place so that further stress and injury are minimized and some seed production can be achieved. The responses to water stress are modified by interaction and integration with other environmental variables, e.g. mineral nutrition, temperature, humidity, age, previous plant history, disease and probably with all external environmental influences; they are not therefore reflexive responses. Clearly decisions are made by the whole plant.
The similarities between avoidance responses in neural circuitry and plant water stress are: (1) a graded response in both cases according to strength of stimulus; (2) similar transduction mechanisms with the different strengths of stimuli; (3) morphological changes in nerve cells and plants induced only by the stronger stimuli; (4) the result of neural learning is to coordinate the behaviour of different muscles to enable an avoidance response by movement. The result of plant learning is to coordinate the developmental behaviour of different tissues to produce an avoidance response by phenotypic plasticity. Muscles are as constrained in their behaviour as any plant tissue, there are just many of them that can be coordinated together to generate great varieties of behaviour. (5) Animal learning lays down additional pathways of communication. Plant learning increases vasculature and increased communication between cells through plasmodesmata (see below). (6) Both organisms integrate the present organismal state to modify the response to further signals. Morphological changes in plants do act like long-term memory, because they will influence subsequent behaviour by the individual plant when other environmental signals are imposed. It can be objected that long-term animal memory is reversible in the absence of further stimulation, whereas morphological changes are not. However, this is not the case. In the short term, stomata usually open again within a few days when water stress is still imposed. ‘Xeromorphic’ leaves are often the first to be abscised after rewatering and new leaves are formed by bud break. There is root turnover and death (Bazzaz, 1996) enabling some recovery of root : shoot ratios.
Do seedlings learn about their environment?
The seedling stage is the most vulnerable for any higher plant, with chaotic fluctuations at the soil surface in temperature, moisture, carbon dioxide, light, patchy nutrient dispersal and the common but variable enemies of disease and predation. There is also a stochastic character to seed dispersal, dormancy breakage, degree of phenotypic individuality (Bradford and Trewavas, 1994) and thus indications that the behaviour of every seed will differ from that of others in certain aspects of behaviour (Bazzaz, 1996). The integrated environment can be viewed as a topological surface continually changing in shape that is directly mapped onto the signal transduction network in sensitive cells and tissues in mirror image, eliciting responses to navigate the environmental maze (Trewavas, 2000). Each seedling must experience a unique spatial and temporal environmental surface. Bazzaz (1996, p. 168) illustrates topological surfaces constructed from the interaction of two environmental variables on different genotypes.
It is recognized that signal transduction mechanisms can be represented as a network. The implication may be that pathways of information flow between the signal and response may not be invariant between different individuals (McAdams and Arkin, 1999; Csete and Doyle, 2002; Elowitz et al., 2002; Guet et al., 2002; Levsky et al., 2002). What is suggested is that when a seedling first receives a signal, a weak response is constructed using the signal transduction constituents to hand and with the signal information finding various channels through which it can flow. Further signalling reinforces this information channel by synthesis of particular signal transduction constituents, much as increased numbers of dendrites improve information flow rates during neural network learning. The signal transduction network thus learns (Trewavas, 2001). Seedlings that fail to learn adequately, quickly die off. It is already known that Ca2+-dependent and -independent processes can be separately invoked to induce identical physiological processes (Allan et al., 1994), and that the synthesis of many constituents concerned with calcium signal transduction are synthesized following signalling (Trewavas, 1999, 2001).