These are my hastily taken and edited notes on the first ALIFE 2020 conference keynote, presented by Mike Levin.
Robot Cancer - why don’t (today’s) robots get cancer?
By Michael Levin, director of the Allen Discovery Center at Tufts and the Tufts Center for Regenerative and Developmental Biology.
Main topics to be covered:
- Biological systems have remarkable structural and functional plasticity that goes beyond genome default outcomes. They feature anatomical homeostasis and top down control of collective outcomes.
- Bioelectric networks is how evolution expanded boundaries of cells into organisms, allowing their goals to get bigger. This has implications for reprogrammability of life
- Hypotensis: multi scale autonomy of goal seeking sub-units, which brings the risks of defection from the global plan (ie cancer), but is overall the key to adaptive robustness. Allows biology to be very reliable on a large scale despite huge amounts of noise and unreliability at the microscale.
Like the brain, somatic tissues form electrical networks that make decisions about anatomy. This proto-cognitive system can be exploited to understand top down programming of cell swarms toward large scale outcomes.
Complex decision making at all levels and scales of biology
A tadpole with no eyes in the head but an eye in the tail can see out of its tail eye perfectly well. The eye in the tail puts out an optic nerve that looks around and finds a spinal cord. It interfaces with the spinal cord to send synapses toward the brain. The brain can dynamically adjust its behavioural programs to accommodate different body architectures.
Memories which caterpillars make survive after their metamorphosis into butterflies. During metamorphosis, the caterpillar’s brain is basically liquefied. It reassembles a brand new brain, yet its old memories remain.
What are implications of this for models of learning in artificial neural networks? In biology, the body can be drastically altered and the information still remains. We need to understand how this works and how biology manages to do this kind of transfer learning. How does it convert caterpillar information to a butterfly architecture?
Another example: planarians. You can take a single flat worm, teach it something simple, chop off the head, take tail fragment which regenerates a new brain. As soon as the new brain is complete, the rest of the body has imprinted the old learning onto it. The memory is stored somewhere else and imprinted onto a brand new brain.
In biology we don’t just deal with a static fixed organism. The body is incredibly plastic. These models raise questions:
- What is a memory?
- How is it encoded, decoded, stored, transferred?
This has both biomedical implications as well as philosophy of mind implications.
What controls the pattern of the body in the first place?
We all start life as a single fertilized egg. As the egg divides, we self-assemble into a very complex machine. All of our organs and tissues have to end up in exactly the right location. Where does this pattern come from? We usually associate DNA with our construction, our “map”, but if you read a genome you see protein sequences - nothing direct about the symmetry, size, shape, organs.
So how do cell groups know what to make? When to stop? And what can we do to convince them to repair themselves? How far can we push the process? If we want these cells to build something totally different, is that possible?
Genes interact in networks. Some networks make proteins, and proteins interact via physics - some are sticky, exert force, etc. What molecules would you tweak at the lowest level of description to give you desired system level outcomes?
Embryogenesis is reliable, but not all hardwired. It has the ability to regulate to specific outcomes.
In adulthood, some animals are able to regenerate large portions of their body. Axolotl can regrow completely or partially their limbs, brain, heart, tail, ovaries, etc. It can regrow a perfect replica of the original and then stop. How does it know when to stop?
A salamander tail grafted onto position of a leg regrows into a leg - it stops being a tail grows toes and other structures associated with a leg. The overall structure knows that even though the tip of a tail is a tail, it needs to be a leg in the larger body plan.
Anatomical set points and regeneration
There is some sort of anatomical set point. You can cut a planarian into pieces (up to 275 pieces as far as we know today), into any direction, and every piece will regrow exactly what is missing until you get a perfect worm. The system somehow knows what a correct worm looks like. There actually no such thing as an old planarian - individual cells will die, but it continuously regenerates them.
This type of regeneration is not just for lower animals. Human liver regenerates; children’s fingertips can regenerate; deer regenerate huge amounts of bone every year.
If you take aggressive carcinoma cells and stick them in a mouse embryo, you will get a normal mouse. The cells will be normalised by the rest of the body.
So: Living anatomy is incredibly plastic. it can repair toward specific anatomical states from different starting conditions. We can observe collective decision making in cell swarms to move towards a common body plan. Mostly we look at stem cells in this process and genomic editing.
Basic questions about genomes and anatomy
In most cases, if we have a mutation in our body our children do not inherit that mutation.Planaria reproduce by fission and regeneration and their body mutations propagate. They have somatic inheritance. Every mutation that does not kill the stem cell ends up propagating and amplifying.
Yet despite 500 million years of somatic inheritance, the regenerative pattern is correct with 100% certainty. How?!
If you take two planaria with different head shapes, transfer some stem cells from one head to another head, then amputate the head - what head shape will regenerate? We do not have any models to make a prediction about this experiment. We do not understand how the collective makes decisions about how the cells decide the correct head shape.
Individual cells and collectivism
Individual cells are very competent. Lacrymaria can handle all of its biological needs at the level of a single cell, such as hunting for bacteria. The cells did not give up their intelligence when they joined into larger, multi-cellular bodies, but they did learn how to collaborate towards much larger goals.
What is a correct frog?
In order to become a frog, tadpoles have to rearrange their face. They have to move their eyes, nostrils, brains, jaws. You must deform to go from tadpole to frog. At one point it was thought that the transition is hardwired - the eye moves to the right by this amount, the brain moves like so, etc etc. But if you move all features of a tadpole on the face, they still become normal frogs. Genetics does not specify hardwired rearrangements: it specifies a system that executes a highly flexible program that can recognize unexpected states and take corrective action, and stop when a perfect frog face has been reached. How does the system know what a correct frog looks like?
Rewriting the goal
Instead of targeting the individual genes and trying to intervene at lowest level of genes and proteins, could we simply rewrite the goal state? Can we reset the set point if there is one? Can we leave the cells alone and let them do what they do best?
This is an important strategy because currently biology is good at figuring out what genes interact with other genes, but what we really want to understand is anatomy - what makes a hand different from a foot? And how can you get from different starting positions to the same outcomes?
The end game is an anatomical compiler. Imagine sitting down in front of CAD-type system, drawing our animal at the level of anatomy. Define what you want, then offload computational complexity onto the cells. Control the outcome with inputs (=experiences), not cellular rewiring.
Back in the day in computer science, to program you had to rewire the machine and interact with it at the level of hardware. It was very tedious. We made incredible progress by realizing that if the hardware is good enough, you don’t need to rewire it to run different software. You can work at a higher level and program with inputs, not by physical rewiring. Can we do the same with anatomy?
How do we go about rewriting anatomical goal states?
In a brain, cells are controlled by synapses. Each cell has ion channels. That is the hardware. The brain software is behaviour and cognition.
The idea is that if we know how to interpret electrical signals, we can know what an animal is thinking about. But it turns out all cells do this, not just brain cells. All cells have ion channels. Most have electrical synapses to their neighbours. We want to be able to read these electrical patterns and figure out what cells are saying to each other in terms of getting closer and closer to their anatomical set points.
Electrical traces in an frog embryo are an early memory of what the face is meant to look like - we can actually see the frog face in the electrical patterns. It then guides the genes to turn on and off to build the physical face. The electrical pattern is a scaffold. If we change the electrical pattern, the features will form in that changed electrical pattern.
So how do they edit electrical patterns?
In the speaker’s lab, they are not using electrical fields. They open and close native ion channels in electrical synapses that control these networks. For example, they take non neural tissue in the body, open and close the synapses to control electrical potential of those cells. This gives control of cell states and propagation of those states to other neighbours. When you do this, you can make radical coherent changes to anatomy.
You can amputate head and tail of a planarian. The middle fragment needs to figure out where head goes and where tail goes. If you manipulate the electrical gradient encoding this pattern, you can build a two headed or no headed worm. We do not know how to build a planarian head from scratch, but we do know the signal that says “build a head right here”. Note that this is not theory, they have already done this.
You can also call up heads belonging to other species. Start with a planarian with a triangular head. Chop off the head. Perturb the network topology, then wash out the drug. The network settles down in a different bioelectric state. Sometimes it settles down to build the same triangular head. But sometimes it will settle down into states corresponding to a flat head or round head. These animals are millions of evolutions distant from a triangular headed flatworm. Yet without touching the DNA at all, you can call up attractors in a state space of the bioelectric circuit which correspond to other species of planarian. The brain and distribution of stem cells also becomes exactly like other species.
In similar ways you can also make spiky flatworms, cylindrical flatworms, weird flatworms that are “life as it could be”. No changes in DNA is required, just in electrical conversations these cells have with each other to figure out what they should be building.
One thing to realize: do not think all we can do is screw up normal patterns. The point is that we now have good computational understanding of where the bioelectric signals come from to induce rational repair. So we can take tadpoles with bad mutations to rationally repair them and get them from a defective state to a totally normal tadpole brain.
Electrical circuits can store memories and patterns that are not the genetic default. We can now directly see the representatoin of large scale goal states and rewrite those memories. You can change electrical pattern and dictate what the animal will do in the future if it gets injured. Eg take a one headed worm. Change its bioelectric pattern. Then cut off the head and tail - it will reform as per new electric pattern.
If you make a two headed worm and weeks later cut off the head and tail again, it will again and again regenerate as a two headed worm. A brief experience of a bioelectric circuit leads to a permanent alteration of the pattern to which the animal regenerates upon damage. The correct pattern is not just rewritable - it is stable. We do now know how to force it back to normal.
These are memories. They are long term stable, rewritable, involve conditional recall (latent until animal is injured), and use the same mechanisms as the brain.
Molecular biologists do not like idea of cells that are not brains having any sort of goal state, but we can see the goal states, and see the set points - and not only see them but rewrite them.
DNA is the hardware
The speaker proposes that DNA is the hardware. It tells you what proteins a cell gets to have; but once you have nailed that down, it forms an excitable medium with some interesting properties. Anatomy is controlled by modular software. The same genetic hardware can house multiple different software outcomes.
Multi-scale organization in artificial life.
All cognitive agents are made of parts; any agent that is intelligent will be made of something. Bodies are just swarms of cells. The magic is not just in molecular mechanisms, but in understanding how these swarms deploy creative problem solving to get you to the same goal in different circumstances.
We are trying to understand how individual cells scale up their computational worlds to tissues, organisms, etc.
“the struggle of the parts” is the idea that the individual component of a biological system is itself goal directed. The large organism may have behavioural goals, but every organ and tissue and cell has goal-directedness. If you put an eye on the tail of a tadpole you get an eye that tries to carry out the function that it has - to pass on visual information to the biggest set of nerves it can find.
The speaker thinks this multi scale autonomy is what is responsible for evolvability. If you have a mutation where your eyes end up elsewhere, you will likely do fine as the eye is competent at forming anywhere and passing its information up to the brain. Competency of individual parts means that if you have a mutation that changes something, all parts around it will still try to execute correct outcomes. Tolerance for mutations becomes massive and makes it much easier to evolve.
Why robots don’t get cancer
Each cell has a boundary between the self and the world. Most of our cells are linked with the rest, recognizing that it is a part of a greater organism. But if you have one cell that becomes electrically isolated from the rest of the body, the separated cell now has a very small radius of “self”. It is no longer plugged in. These cells try to proliferate as much as they can and go wherever they want, as all unicellular organisms do. These cancer cells are not more selfish than any other cell type, they just have smaller “self"s. Cancer cells have very tiny goals and treat the rest of the body as environment.
In robotics, machines don’t really have individual parts that run off and do their own thing. This is why robots don’t get cancer.
In cells, if there is a tumour you can artificially force those cells into communication with their owner. These cells do not need to be killed, just to be plugged into the rest of the communication networks.
- Computational capabilities, memory are ancient and not just for brains. All cells have this, brains came from this.
- Evolution exploits physics of bioelectricity to implement networks that store large scale patterns which serve as memories.
- Scaling of allostatic setpoints enables the cooperation of smaller, competent parts into remarkably plastic and robust functional forms
- All of this contains consequences for different fields:
- Bots and AI: new, non neuromorphic approaches to machine learning
- Biomedicine: exploit intelligence of cell collectives by using stimuli to reset their goal encodings, not rewiring their hardware/DNA
- Artificial life: combine bottom up emergence with top down cognition