Here is something that took me years to fully absorb, even though I encountered it in my first week of graduate school: if you take two genetically identical bacterial cells, put them in the same growth medium, at the same temperature, with the same nutrients, they will not do the same thing. One will divide in 42 minutes. The other will divide in 58 minutes. One will express a particular gene at twice the level of the other. One will survive an antibiotic pulse that kills its twin.
Same genome. Same environment. Different outcomes.
This is not noise. This is not experimental error. This is biology. The molecular machinery inside a cell, the transcription factors searching for binding sites, the ribosomes translating mRNA, the enzymes catalyzing reactions, all operate at concentrations so low and in spaces so small that thermal fluctuations, the random jostling of molecules in solution, produce measurable differences in behavior from one cell to the next, from one moment to the next. Gene expression is not a deterministic program. It is a stochastic process where the same input produces a distribution of outputs, and that distribution is not a bug. It is a feature. It is how populations hedge their bets against uncertain environments. It is how antibiotic resistance emerges. It is how tumors evade chemotherapy. It is how life persists.
And yet. The dominant paradigm of biological discovery, the one most of us were trained in, treats biology as though it were fundamentally deterministic: perturb the system, measure the average response, build a mechanistic model that explains the average, declare the mechanism understood. The entire statistical framework of biological research, hypothesis testing, p-values, fold changes, is built to extract a single summary number from a distribution and then discard the distribution. We literally throw away the stochasticity and keep the mean. Then we are surprised when our predictions fail in individual cells, individual patients, individual organisms.
I want to argue that this paradigm is reaching its limits, and that the convergence of AI, large-scale biological data, and computational infrastructure creates the conditions for a fundamentally different approach to discovery. Not a better version of what we have been doing. A different thing.
The Three Properties of Biology That Break Deterministic Thinking
Before proposing what comes next, I want to be precise about what makes biology different from the systems that deterministic science was designed to study. There are three properties that, taken together, make biological systems fundamentally resistant to the kind of reductionist analysis that works so well in physics and chemistry.
Stochasticity: Same Input, Different Output
At the molecular level, biological processes are governed by stochastic kinetics. Transcription initiation is a probabilistic event: a transcription factor must find its binding site by diffusing through the nuclear volume, and the time it takes to do so varies randomly from one event to the next. Translation is noisy: the number of protein molecules produced from a single mRNA varies. Enzymatic reactions depend on random encounters between enzyme and substrate molecules. At the single-cell level, these molecular fluctuations compound into measurable phenotypic variation: different growth rates, different stress responses, different decisions about whether to differentiate, divide, or die.
This stochasticity is not something biology tolerates. It is something biology exploits. Bet-hedging strategies, where a genetically identical population produces phenotypic diversity to survive unpredictable environments, are found across the tree of life. Bacterial persistence, where a small fraction of cells enter a dormant state that survives antibiotic treatment, is a stochastic phenomenon. Cancer heterogeneity, where tumor cells within the same tumor have different drug sensitivities, is partly driven by stochastic gene expression. The randomness is doing work.
Redundancy: Many Paths to the Same Outcome
Biological systems are pervasively redundant. Multiple genes perform overlapping functions. Multiple metabolic pathways can synthesize the same essential metabolite. Multiple signaling cascades converge on the same downstream target. Multiple regulatory mechanisms control the same gene. This redundancy means that knocking out a single gene often has no observable phenotype, not because the gene is unimportant but because the cell reroutes through an alternative path. The system is robust to perturbation precisely because it has built-in alternatives.
This creates a fundamental problem for reductionist analysis. The standard experimental approach in molecular biology is to perturb one variable (knock out a gene, inhibit an enzyme, add a drug) and measure the response. But in a redundant system, the absence of a response does not mean the variable is unimportant. It means the system compensated. And the compensation itself is biologically meaningful: it reveals the architecture of the network, the hierarchy of backup systems, and the priorities the cell places on different functions. But that information is invisible to a paradigm that looks for single-variable effects.
I encountered this directly in my metabolic modeling work. When I built genome-scale models for Halomonas elongata, flux balance analysis would predict that certain gene knockouts should be lethal because they remove the only path to an essential metabolite. In the actual organism, the knockout was viable because the cell activated an alternative pathway that the model did not include, or that was annotated as belonging to a different metabolic subsystem entirely. The model was not wrong about the stoichiometry. It was wrong about the network architecture because the redundancy was not captured in the annotation.
Context-Dependence: Function Is Not Fixed
The same gene does different things in different cells, at different times, under different conditions. A transcription factor that activates a growth pathway in one cellular context represses a stress response pathway in another. An enzyme that catalyzes one reaction in the cytoplasm catalyzes a different reaction when localized to the membrane. A signaling molecule that promotes cell division at low concentrations promotes cell death at high concentrations.
This context-dependence means that the concept of "gene function," as a fixed property of a DNA sequence, is an approximation that breaks down at the resolution where most interesting biology operates. Function is a relationship between a molecular entity and its context, not a property of the entity alone. This is why gene ontology annotations, which assign fixed functional categories to genes, are both indispensable and fundamentally incomplete. They capture the average function across the contexts in which the gene has been studied, which may not include the context you care about.
Biology is not a machine that executes a program. It is a population of molecules that negotiate outcomes probabilistically, within a network of redundant alternatives, in a context that changes constantly. Any paradigm of discovery that does not account for all three of these properties will systematically miss what matters most.
What Computation Actually Does in a Stochastic Science
If biology is non-deterministic, what is the role of computation? This question matters because it determines what we build, what we measure, and what we consider a successful analysis.
In a deterministic science like classical mechanics, computation's role is straightforward: solve the equations that describe the system, predict the outcome, verify with experiment. The computation replaces hand calculation. The logic is the same. You have a model, initial conditions, and a prediction.
In a stochastic science, computation plays a fundamentally different role. It does not predict specific outcomes. It characterizes distributions of outcomes. It does not identify the mechanism. It identifies the ensemble of mechanisms that are consistent with the observed data. It does not confirm a hypothesis. It quantifies the relative plausibility of competing hypotheses given the evidence.
This shift, from prediction to distribution, from mechanism to ensemble, from confirmation to quantification of uncertainty, is not just a technical adjustment. It changes what counts as understanding. In the deterministic paradigm, understanding means being able to predict exactly what will happen. In the stochastic paradigm, understanding means knowing the shape of the distribution, knowing what controls its width, knowing which variables shift its center, and knowing when the distribution itself changes shape in response to perturbation.
The Specific Roles of Computation in Stochastic Biology
Computation in this framework serves four distinct functions, and conflating them leads to confusion about what a computational analysis has actually accomplished.
Pattern detection: finding structure in high-dimensional, noisy biological data that human perception cannot access. This is where dimensionality reduction, clustering, and unsupervised learning operate. The computation identifies groups, gradients, outliers, and transitions in data spaces with thousands or millions of dimensions. It does not explain why those patterns exist. It tells you where to look.
Hypothesis generation: using detected patterns to propose mechanistic explanations that can be tested experimentally. A clustering analysis that reveals two distinct subpopulations in what was thought to be a homogeneous culture generates the hypothesis that there are two distinct cellular states, which can be tested with single-cell experiments. The computation does not prove the hypothesis. It generates it.
Quantitative constraint: using mathematical models to determine which mechanistic explanations are consistent with the observed data and which are ruled out. A metabolic model that shows a particular flux distribution is thermodynamically infeasible eliminates that explanation. A statistical model that shows a particular gene's expression pattern is inconsistent with a proposed regulatory mechanism constrains the space of viable explanations. The computation narrows the field.
Distributional prediction: predicting not what will happen, but the probability distribution of what might happen. Stochastic simulation algorithms, Bayesian inference, and probabilistic graphical models all operate in this mode. They do not produce a single answer. They produce a distribution of answers, with associated uncertainties, conditioned on assumptions that must be stated explicitly. This is the most honest form of biological prediction, and the one most aligned with the actual nature of the system being studied.
Why the Current Discovery Paradigm Is Approaching Its Limits
The paradigm that has dominated biological research for a century follows a linear logic: observe a phenomenon, form a hypothesis about its mechanism, design an experiment that tests one variable, collect data, apply a statistical test, accept or reject the hypothesis, publish, repeat. This is the hypothetico-deductive method, inherited from physics, adapted for biology. It has been extraordinarily productive. It gave us the central dogma, the genetic code, the structure of DNA, the mechanisms of cell division, and the molecular basis of disease.
But it has a structural limitation that becomes more acute as biology moves toward systems-level, multi-omic, high-dimensional investigation: it was designed for low-dimensional, single-variable problems. When you are asking "does gene X affect phenotype Y," the hypothetico-deductive method works beautifully. When you are asking "how do 5,000 genes, 2,000 metabolites, 10,000 regulatory interactions, and stochastic molecular dynamics collectively produce a phenotype that varies across individual cells and changes over time," the method breaks. Not because it is wrong, but because the hypothesis space is too large, the variables are too interconnected, and the stochasticity makes single-experiment conclusions unreliable.
The result is a growing disconnect between the data we can generate and the questions we can answer. We can measure the expression of every gene in every cell in a tissue. We can sequence every genome in a microbial community. We can profile thousands of proteins in a single experiment. But our framework for converting that data into understanding is still organized around testing one hypothesis at a time, about one variable at a time, using statistical methods designed for small, clean, low-dimensional datasets.
A New Paradigm: Stochastic Exploration at Scale
What I am about to describe is not a finished framework. It is a direction of travel that I believe the convergence of AI, large-scale biological data, and computational infrastructure makes possible for the first time. I am calling it stochastic exploration at scale, not because the name is elegant, but because it captures the two essential features: the exploration is probabilistic rather than deterministic, and it operates across the full complexity of biological systems rather than one variable at a time.
Principle 1: Replace Hypothesis-First with Exploration-First
The traditional paradigm starts with a hypothesis and designs an experiment to test it. The new paradigm starts with a comprehensive measurement and uses computation to discover which hypotheses the data supports. This is not hypothesis-free science. It is hypothesis-generating science, where the hypotheses emerge from the data rather than preceding it.
Single-cell RNA sequencing is already operating in this mode, even if most practitioners do not frame it this way. When you profile 50,000 cells from a tumor and perform unsupervised clustering, you are not testing a hypothesis about cell types. You are asking the data to reveal structure that you did not anticipate. The clusters you discover, the trajectories you infer, the gene programs you identify, these are computationally generated hypotheses about the biological organization of the tissue. The experiment that follows, a perturbation that targets a specific cell state, or a spatial measurement that validates a predicted interaction, is the test. The computation comes first. The experiment comes second. This inversion of the traditional order is not a loss of rigor. It is an adaptation of rigor to the complexity of the system.
Principle 2: Model Distributions, Not Averages
The new paradigm treats biological variability as signal, not noise. Instead of computing the mean expression of a gene across a population and discarding the variance, it models the full distribution and asks: what controls the shape of this distribution? What makes it bimodal? What makes it heavy-tailed? What makes it shift under perturbation?
This is where AI offers something genuinely new. Foundation models trained on millions of single-cell transcriptomic profiles learn representations that capture not just the average cell state but the landscape of possible states and the transitions between them. A model like scGPT does not predict what a single cell will do. It learns the probability distribution over cell states and how that distribution responds to perturbation. This is a fundamentally different kind of biological knowledge than what a differential expression analysis provides. It is distributional knowledge, and it is the kind of knowledge that matches the stochastic nature of the system.
Principle 3: Embrace Redundancy Through Combinatorial Perturbation
If single-gene perturbations produce no phenotype because the system compensates, the solution is not to give up on perturbation. It is to perturb combinatorially and at scale. CRISPR screens that knock out two, three, or four genes simultaneously can reveal synthetic lethal interactions, epistatic relationships, and redundancy architectures that single perturbations miss entirely. The combinatorial space is vast, far too large for exhaustive experimental coverage, but AI systems can explore it intelligently.
This is where AI agents become essential, not as replacements for human judgment but as navigators of combinatorial spaces that humans cannot traverse manually. An agent that can analyze the results of a first-round perturbation screen, identify the most informative combinatorial perturbations to test next, design the experiment, and interpret the results is performing a kind of adaptive exploration that the traditional paradigm cannot support. The agent does not need to understand the biology in the way a scientist does. It needs to identify which experiments will maximally reduce uncertainty about the system's architecture. That is an optimization problem, and AI is very good at optimization problems.
Principle 4: Build Probabilistic, Not Deterministic, Models
The genome-scale metabolic model I build using flux balance analysis produces a single optimal flux distribution. That is a deterministic prediction about a stochastic system. A more honest model would produce a distribution of flux states, weighted by their probability given the available data and the known constraints, with explicit uncertainty on every prediction.
Bayesian approaches to metabolic modeling do this, but they have historically been computationally expensive and difficult to scale. Foundation models change the economics. A foundation model pretrained on thousands of metabolic datasets across hundreds of organisms learns a prior distribution over metabolic behaviors. When you condition that model on data from a specific organism under specific conditions, it produces a posterior distribution: a probabilistic prediction that reflects both the general patterns learned from the training data and the specific evidence from the organism at hand. This is not just statistically more rigorous than a point prediction. It is biologically more honest.
Principle 5: Close the Loop Between Computation and Experiment
The most important structural change in the new paradigm is the elimination of the gap between computation and experiment. In the traditional paradigm, computation and experimentation are sequential: you analyze data, then you design the next experiment, then you wait for results, then you analyze again. Each cycle takes weeks to months. The lag between analysis and action is where momentum dies and context is lost.
The new paradigm operates in a tight loop. Computational models generate predictions continuously. Experimental platforms test those predictions in near-real time. Results flow back into the model immediately. The model updates, generates new predictions, and the cycle continues. This is the lab-in-the-loop concept, and it transforms discovery from a discrete, step-by-step process into a continuous, adaptive one.
The key enablers are automation (robotic experimental platforms that can execute experiments designed by algorithms), foundation models (that can predict experimental outcomes before they are measured, prioritizing the most informative experiments), and AI agents (that can orchestrate the entire cycle without human intervention for routine decisions, escalating to human judgment only for genuinely novel or ambiguous results).
The new paradigm does not replace the scientist. It changes what the scientist does. Instead of manually executing the discovery cycle, the scientist defines the question, evaluates the quality of the automated exploration, and exercises judgment on the results that matter. The scientist becomes the curator of an intelligent discovery process, not the laborer within it.
What This Looks Like Concretely
Let me make this tangible with an example from my own domain: metabolic engineering of microbial strains for bioproduct synthesis.
In the current paradigm, we work like this. We build a genome-scale metabolic model. We run FBA to identify gene knockout targets. We engineer the strain. We grow it. We measure the product titer. If the titer is low, we hypothesize about what went wrong (the model missed a regulatory interaction, the knocked-out gene had an unexpected moonlighting function, the cell compensated through a redundant pathway), and we design the next experiment to test that hypothesis. Each cycle takes three to six months. A PhD thesis covers three to five cycles.
In the new paradigm, the same project would work differently. A foundation model pretrained on metabolic data across thousands of organisms provides an initial probabilistic model of the target organism's metabolism, with uncertainty estimates on every flux prediction. An AI agent designs a first-round combinatorial CRISPR screen targeting the 50 most promising knockout combinations, optimized not for expected product titer but for maximum information gain about the metabolic network's architecture. The screen runs on an automated platform. The sequencing data flows through an automated NGS pipeline. The agent analyzes the results, updates the metabolic model with the new data (reducing uncertainty on the pathways that were perturbed), and designs the second-round screen. Within weeks, not months, the model has been refined by multiple rounds of experimental feedback, the most promising engineering strategies have been identified, and the uncertainty on the final prediction is explicitly quantified.
The scientist's role in this process is not diminished. It is elevated. Instead of spending months doing the mechanical work of strain construction, growth measurement, and data analysis, the scientist focuses on the questions that require human judgment: Is the model capturing the right biology? Are the agent's experimental designs asking the right questions? Are there biological factors (regulation, stress responses, evolutionary dynamics) that the automated system is not accounting for? The scientist becomes the quality controller of the exploration, not the engine of it.
The Frontier: Where Discovery Meets Engineering Meets AI
Mapping the Dark Proteome
Between 30% and 50% of genes in most microbial genomes have no functional annotation. These genes encode proteins that do something, but we do not know what. In the current paradigm, characterizing each of these proteins requires years of experimental work per gene. In the new paradigm, foundation models predict probable functions from sequence and structural features, AI agents design high-throughput perturbation experiments to validate those predictions at scale, and the results feed back into the model to refine predictions for the next batch of uncharacterized genes. The dark proteome becomes accessible not one gene at a time, but in parallel, across entire genomes, across entire microbial communities.
Engineering Microbial Communities, Not Just Organisms
The most efficient bioproduction systems in nature are communities, not monocultures. But designing synthetic microbial consortia requires understanding how species exchange metabolites, compete for resources, and regulate each other's behavior. The interaction space is combinatorial and stochastic: the same two species can cooperate or compete depending on nutrient concentrations, cell ratios, and spatial organization. A discovery paradigm that can explore this space requires high-throughput community cultivation, metagenomics and metaproteomics at community resolution, probabilistic models of interspecies interaction, and AI agents that design the next community composition based on what was learned from the last one. No human can navigate this space manually. It is too large, too stochastic, and too context-dependent. But an AI-augmented discovery system can.
Personalized Biological Models
If biology is stochastic and context-dependent, then the model that predicts how a drug will affect patient A may not predict how it will affect patient B. The new paradigm makes personalized biological models feasible: foundation models pretrained on population-scale data, conditioned on an individual's genomic, transcriptomic, proteomic, and metabolomic profile, producing patient-specific predictions with quantified uncertainty. This is the path to precision medicine that actually works, not because we have eliminated biological variability but because we have built computational systems that can reason about it.
Evolutionary Design
Evolution is the original stochastic optimization algorithm: it generates random variation, selects for fitness, and iterates. It is also breathtakingly slow. AI-guided directed evolution compresses this process: instead of relying on random mutagenesis and screening, foundation models predict which mutations are most likely to improve function, AI agents design libraries that explore the most promising regions of sequence space, and high-throughput screening validates the predictions. The result is evolution at AI speed: the same exploration of sequence space that nature accomplishes in millennia, accomplished in weeks. Profluent's OpenCRISPR-1, a gene editor designed entirely by AI, is an early example of this approach. The paradigm extends to any protein engineering challenge where the goal is to navigate a vast, stochastic fitness landscape toward a functional optimum.
What This Demands of Us
If this paradigm is coming, and I believe it is, it demands a different kind of scientist than the one our training programs currently produce.
It demands scientists who are comfortable with uncertainty, who can reason about distributions rather than point estimates, who understand that "I do not know, but here is my uncertainty quantified" is a more useful answer than a false precision.
It demands scientists who can work with AI systems as collaborators, not just as tools. Knowing how to call an API is not enough. You need to understand what the model learned, what it did not learn, where its predictions are reliable and where they are extrapolating beyond its training distribution. You need to be the quality controller, the one who catches when the AI is confidently wrong about biology.
It demands scientists who can think across scales and across disciplines. The new paradigm integrates molecular biology, genetics, statistics, software engineering, machine learning, and experimental design into a single workflow. No one person can be expert in all of these. But the people who lead discovery teams need to be fluent enough in each to make sound architectural decisions about how the pieces fit together.
And it demands scientists who can hold two things in their heads at the same time: the power of the new tools and the humility to remember what they do not capture. An AI that predicts protein function from sequence does not understand protein function. It has learned statistical regularities in a training set. Those regularities are powerful and useful, but they are not understanding. Understanding requires connecting the prediction to mechanism, testing it against experiment, and knowing when to trust it and when to doubt it. That is judgment. And judgment, for the foreseeable future, remains a human contribution.
The age of AI does not make the scientist obsolete. It makes a different kind of scientist essential: one who can orchestrate intelligent systems, interrogate their outputs, and exercise biological judgment at the frontier where data ends and discovery begins.
Where I Stand
I have built genome-scale metabolic models by hand, reaction by reaction. I have fit kinetic parameters to noisy growth curves and stared at residuals trying to decide whether the misfit was noise or biology. I have assembled metagenomes from mixed communities where the species boundaries were uncertain and the annotations were incomplete. I have trained a transformer on protein sequences and watched it learn patterns I could not have specified in advance.
Every one of these experiences taught me the same lesson: biology resists the frameworks we impose on it. It is noisier than our statistics assume, more redundant than our models capture, and more context-dependent than our annotations record. The frameworks are not useless. They are essential. But they are approximations, and the quality of our science depends on knowing where the approximation breaks down.
The paradigm I have described in this post, stochastic exploration at scale, augmented by foundation models, orchestrated by AI agents, grounded by experimental validation, is not a rejection of the scientific method. It is an evolution of the scientific method, adapted for systems that are too complex, too stochastic, and too high-dimensional for the hypothesis-first, one-variable-at-a-time approach to handle alone.
We are not there yet. The foundation models are still being built. The agent frameworks are nascent. The experimental automation is expensive and unevenly distributed. The data bottlenecks I described in my previous writing, kinetic data sparsity, fluxomics cost, annotation gaps, publication bias toward positive results, are all still real.
But the trajectory is clear. And the scientists who will shape this transition are the ones who understand both the biology and the computation, who can build the bridges between wet lab and algorithm, between stochastic measurement and probabilistic model, between human judgment and machine intelligence.
That is the work. I am building toward it. And I am looking for others who want to build toward it too.
All writing