Louis Scheffer of @HHMIJanelia promises us a whole simulated Drosophila within 5-10 years. If they fail I get beer, if they succeed I buy beer. I am pretty confident. Why do people believe the path from connectomics to simulation is short?
Fruit flies are pretty complicated! They have about 135,000 neurons, lots of connections, and a complex body. This group plans to simulate the entire body -- including the ability to fly, have sex, and so forth -- in a complex virtual environment, within the next 5 to 10 years! Given how much computing power people are willing to devote to large projects like this, I don't think the compute will be the limiting factor.
If they succeed in doing this, then super-accurate mouse models probably wouldn't be more than 5 or 10 years further off. I say this not only because fruit flies are not too far from mice in complexity (they are pretty far, but not that far), but also because the success of such a project would spur on orders of magnitude more time, money and energy (people) devoted to animal simulations.
Other groups have attempted to simulate a C. elegans brain and body:
We can’t simulate all behaviors but can reproduce various locomotion neural dynamics in response to stimuli like fwd, backward, turns and chemo sensation. With current comp tools in few years we’ll be able to simulate #celegans fully. So 10 years for the fly could be doable.
I recall that groups have attempted to simulate a bee brain, which has 1 million neurons -- but have not attempted to simulate the whole body and environment.
Google has some researchers working on Meso-scale simulations, that leverage Deep Learning:
And various other Deep Learning-based projects aim to simulate at least parts of complex animal brains. Here are two related talks that appeared at the COSYNE 2018 conference:
A novel deep recurrent network for predicting large scale population responses to natural video
To understand the representations in visual cortex, we need to be able to faithfully predict neural activity in response
to its natural input: a continuous video stream. Since cortical activity is highly variable and context dependent,
this prediction is already difficult for integrated neural activity to static natural images, and even more
difficult for dynamic responses to movies. In awake animals under free-viewing conditions, eye movements and
brain states add to this response variability, making the prediction problem even harder. While deep convolutional
networks have recently been shown to improve prediction performance over linear-nonlinear type models and
are currently considered state-of-the-art, they make suboptimal use of the data, because they cannot account for
Here, we developed a new deep recurrent network architecture that predicts the deconvolved Ca++ activity of
thousands of simultaneously recorded neurons in mouse V1 to natural videos, recorded at 7Hz and 30Hz, respectively,
while simultaneously estimating dynamic gaze position and brain state changes related to running
state and pupil dilation. In addition to the natural movie input, the network uses pupil position and dilation extracted
from a video of the animal’s eye, as well as treadmill velocity. The unknown relation between pupil position
and gaze position on the monitor is learned by the network during training based solely on predicting neural activity.
We find that incorporating all these elements (nonlinear recurrent network, running speed, pupil position, and
pupil dilation) significantly increases the prediction performance of the network. Our network achieves between
40% and 60% of a leave-one-out estimate of single-trial correlation with the mean response over repeated presentations.
To the best of our knowledge, this makes our model the state-of-the-art on single trial prediction of
dynamic responses to natural movies on large neuronal populations.
A modular neural network model of the primate grasping circuit
Grasping objects is an essential part of primate behavior. In macaque monkeys, the core of the grasping circuit is
formed by the interconnected anterior intraparietal area (AIP), the hand area (F5) of the ventral premotor cortex,
and the hand area of the motor cortex (M1). Generating appropriate delayed grasping movements involves many
inter-related steps, from identification of visual target identity and spatial location, to the determination and maintenance
of the appropriate movement plan, and finally the control of muscles. We hypothesized that the grasping
circuit could be effectively modeled by training a modular recurrent neural network on visual object features to
output muscle dynamics. To train and test our model, we recorded from neural populations simultaneously from
AIP, F5, and M1 using floating microelectrode arrays while two macaque monkeys performed a delayed grasping
task in which ~50 objects of distinct shape, size, and orientation had to be grasped and lifted. During every
trial, arm and hand kinematics were recorded and transformed into a 50-dimension muscle length space using
a musculoskeletal model. The network model was successfully trained to produce single-trial muscle velocities
during grasping (normalized error: <5 %). Interestingly, the internal dynamics of the model matched the recorded
neural data (canonical correlation, mean r=0.7 over 12 dimensions). Furthermore, biological regularizations were
implemented to encourage simplistic solutions, which resulted in a strong alignment between the contributions
of modules of the model and the recorded brain areas to the canonical variables (r=0.80) that was not present
in untrained networks (r=-0.06). Our model therefore provides a simplistic and accurate representation of the
primate grasping circuit and suggests that the combined processing of these areas can be well understood as a
network optimized to transform object information into the muscle dynamics required to grasp each object.
That poster can be downloaded here:
An even weaker sort of simulation leverages Machine Learning, but only uses behavioral data to train it -- in other words, no brain scans; just videos of how the animal behaves. The example of this most people on this forum have probably heard about is this work from AllenAI on simulating dog behavior:
About a year earlier there was this work at ICLR on hierarchical behavioral modelling:
And there was this earlier work on generating videos of mice running around in a habitat (in lab).
In the not-too-distant future, I think the approach that will win out will be meso-scale "simulations" (not really simulations), based on Deep Learning applied to brain population responses + behavioral data (e.g. videos of how the animal behaves), along with neuromuscular models of the organisms interacting with a simulated environment. This seems to be a good compromise between the super-detailed simulations at the individual neuron or synapse level, and the very crude, purely behavior-based models (no brain data).