It turns out the wind-up toy is a good analogy of an organism's biology -- the clockwork that comprises and moves the toy are akin to the genes and the genes' mRNA and protein products of an individual.  In the course of normal everyday living, biology follows a fairly predictable path.  Some genes are constantly on like the turning of gears; other genes turn on and off like the rise and fall of the toy's feet.  Biology can start to go awry due to perturbations such as exposure to toxins and pathogens, just as the toy's position and orientation is potentially changed by obstacles in its path.  A big enough toxin exposure kills the organism, just as a big enough obstacle topples the toy irreversibly.  Additionally, the obstacles and perturbations that cause gross disruption can be internal -- such as loose cogs and cancer.

           A corollary of the similarity between clockwork and biology, if you accept the premise, is that the future trajectory of the organism or toy are predictable, if you're given a movie of its recent past and its current state.  Importantly, predicting the future states does not require you to fully know or understand the internal gears and mechanisms (in the case of the clockwork toy) or the gene regulatory networks (in the case of a living organism).  Next token prediction in the absence of full mechanistic interpretability happens to be what modern large language model (LLM) AI excel at:  Given an input prompt, the set of tokens provided by the user, the LLM will generate additional tokens one at a time, with each new generated token serving as part of the input for generating the next token.  We believe that we can do the same for biology.

          To do this, however, we need enough information to analyze and train on.  We're not going to have much luck predicting whether the toy topples based only on an instantaneous gyroscope reading.  We will need reasonably high-resolution images of the entire toy and the obstacle.  

           The analogous information required for predicting biology forward is gene expression data.  Humans and rats both have about 25,000 genes, and at any given moment for a specific individual, each gene's expression level can be different, varying up to 7 logs.  We can almost think of it like a 160 x 160 pixel grayscale image, with each pixel having an intensity property corresponding to the log gene expression level.  If we have enough gene expression "pictures" of an individual taken with regular timing, then we may be able to reasonably generate the pictures corresponding to the next few days.

           The problem is that this type of data has not really been collected before.  Because RNA sequencing (RNAseq) is expensive, researchers typically will only take a picture before the patient or animal model is dosed with a drug, and one more picture after dosing.  This is like taking a picture of the clockwork toy before it steps onto the obstacle, and after it either topples or recovers.  From this pair of pictures, you can determine whether the obstacle toppled the toy, but you can't figure out how the toy moves, topples, or recovers.