The motivation: simulation as infrastructure for a bio-industrial age
We are pursuing the challenge of a new industrial age in which sustainability is not a constraint bolted onto chemical manufacturing but the consequence of switching its basis: from petrochemistry and high-temperature catalysis to biological processes that run at ambient conditions on renewable inputs. Enzymes replace heavy-metal catalysts; engineered cells replace reactors; grown and self-assembled materials replace energy-intensive fabrication; biologics and precision therapeutics replace blunt small molecules. At present, empiricism sets the pace of this transition: directed evolution, screening, and formulation are slow, costly, and largely unpredictive. A general-purpose biomolecular simulator that could predict rather than screen would be the enabling infrastructure of that age, the equivalent of computational modelling for aerospace or semiconductors. However the simulation stack we have was built to answer the question posed by structural biology, “what is this molecule and how does it bind” , whereas the bio-industrial age asks a different one: ”what does this molecule do, in context, over time, and can we make it and trust it. ”
The cross-cutting challenge
Almost every high-value gap shares one root: today’s tools assume dilute conditions, a single equilibrium structure, fixed chemistry, and reachable timescales. The bio-industrial questions violate all four. The grand challenge is to reorient simulation from binding in a box to consequence in a body, a cell, or a material, over time. Such simulation must be non-equilibrium, kinetic, reactive, and multiscale by default.
Therapeutics
The dominant challenge is not affinity, which the computational chemistry industry already pursues, but molecular fate. Specifically the challenge is to predict how the body chemically transforms a drug (metabolism), into what, and how fast, at a throughput that informs design. This is reactive enzyme chemistry (cytochrome P450 oxidation), historically blocked by the cost of per-compound QM/MM. Reaction-specific machine-learned potentials are the credible route, because they move the burden of generalisation from an effectively infinite drug-space to the small, finite repertoire of metabolic reaction chemistries. The remaining wall, and one that is rarely discussed, is the trustworthiness of the reference quantum chemistry for iron-oxo catalysis. Alongside this sit kinetics and residence time (rather than equilibrium affinity) and proteome-scale selectivity as a safety check at design time.
Industrial biotechnology
Here the bottleneck is manufacturability, not function: whether an engineered enzyme or protein will express, fold in vivo, avoid aggregation, and work in the crowded, non-equilibrium cellular context rather than a dilute simulation box. A related grand challenge is predictive, high-throughput enzyme catalysis: ranking mutations by their effect on catalytic rate (a transition-state problem) before wet-lab screening. This is already pursued but remains unsolved at the accuracy and throughput that design requires. Cracking it would displace the largest empirical cost in biocatalysis.
Biomaterials
Here the material is its processing history. Properties emerge from non-equilibrium assembly, from evolving chemistry (crosslinking, self-healing, degradation, where network topology is the dependent variable), and from the mesoscale between atoms and continuum. The challenges that remain largely unsolved are simulating assembly and processing themselves, modelling reactive materials whose bonds form and break during the simulation, and failure and lifetime under load, not merely the equilibrium native state. Coarse-graining and mesoscale methods have established communities, but a principled bridge across scales that preserves the chemistry remains out of reach.
The delivery approach
A simulator this broad cannot be built as a monolith, and should not be. The tractable path is component by component: each module tackles a specific, well-posed question that sits inside one of the grand challenges ( a reaction-specific potential for metabolic oxidation, an in-cell crowding environment, an assembly-and-processing engine, a mutation-to-rate estimator), and is built and validated on its own terms before being combined with the others. This mirrors how the challenges themselves decompose: they are not one problem but a family of physics, and each yields to a focused effort where a general assault would stall.
Each component should be released as professionally maintained open-source software: not as a code dump accompanying a paper, but as durable, documented, tested infrastructure with real maintenance behind it. This matters for three reasons. Firstly, hardening: a component only becomes trustworthy when a community can stress it against problems its authors never imagined, and open, maintained code is the only mechanism that generates that adversarial testing at scale. Secondly, composability: openly specified, well-engineered modules can be built on and combined by others, so the simulator grows through contribution rather than solely through one organisation’s effort. Thirdly, trust: bio-industrial and clinical decisions cannot rest on black boxes, and open methods can be audited in a way that proprietary tools cannot.
Written by Julien Michel and proofed by Claude.