DNA May Not Need to Be Learned Before It Can Be Calculated

There is a tradeoff buried underneath modern computational biology. If you want deep physics, computation becomes expensive. If you want speed, you usually move toward approximations, parameterized models, coarse-graining, or machine learning.

That tradeoff has shaped an enormous amount of how we compute biological structure.

What if it is not fundamental?

That is the question behind Flux Genome Physics. The first answer is now executable: Flux can construct a complete backbone-first B-DNA duplex from the same physics engine we use for atoms, molecules and bonds — without a DNA-specific fitted force field or a machine-learned structure model.

The measured helical rise is not supplied as the target. It comes out at 3.16 Å, compared with 3.40 ± 0.32 Å across the primary deposited B-DNA reference. And the mechanism is the part that matters most: base stacking alone does not set the rise. Closing the physical sugar-phosphate backbone creates the equilibrium.

From a roadmap to an executable DNA model

When we published the Genome Physics program in July, we framed it around a simple principle:

A DNA sequence tells you which physical constituents are present. But what that sequence can actually do depends on pairing, stacking, backbone geometry, mechanics, environment, interactions, modifications, and structural state.

The roadmap therefore started deliberately at the bottom: before mutations, oligo design, regulatory sequences or inverse search.

The first requirement was a Nucleotide and Duplex Core: bases, pairing, stacking, nucleotide geometry and the physical backbone built on the same Matter Computing foundation already used for atoms, molecules and interactions.

That core now exists as an active research module.

Building B-DNA from the physics upward

This is not a statistical model that receives thousands of DNA structures and learns what a helix normally looks like. It is not an AMBER or CHARMM force field fitted specifically to reproduce nucleic-acid behavior. And it is not a DFT calculation of the electronic wavefunction.

The same underlying Flux bond-length, contact, hydrogen-bond, geometry and field machinery used elsewhere in the platform is being extended upward into nucleic-acid matter.

The last row is the one that matters most. Because the 3.16 Å value was not supplied as the answer.

The backbone sets the rise

A DNA double helix advances along its axis by a characteristic distance from one base pair to the next. That distance is the helical rise. Across the 286 dinucleotide steps in the primary deposited reference (1EQZ), the mean is 3.40 ± 0.32 Å.

So what determines it?

An intuitive answer might be base stacking. The bases attract and stack. Perhaps their preferred contact distance simply determines the spacing.

That is not what happens in the current Flux calculation. Release the bases under the stacking interaction alone and they continue pulling inward. The energy does not find a stable minimum near the biological rise. It keeps moving down toward approximately 2.91 Å.

Then the sugar-phosphate backbone is physically closed. The O3′–P–O5′ phosphodiester linkage now has to remain compatible with the base stack, the nucleotide geometry and the rest of the backbone. The problem changes. What was previously a one-sided inward attraction becomes a bounded two-sided physical well.

In other words, in this calculation:

The bases create the attraction. The backbone creates the equilibrium.

That is more interesting than simply landing near the right number. It is a mechanism. The rise appears because two parts of the physical system that want different things have to satisfy one another simultaneously.

The backbone sets the rise.

Why “no fitted parameters” matters here

There is an important distinction to make. DNA simulation itself is not new. Neither is physics-based DNA computation.

Modern molecular dynamics can reproduce extraordinary amounts of nucleic-acid behavior. But its success depends on carefully developed nucleic-acid force fields. Parmbsc1, for example, was explicitly parameterized using high-level quantum-mechanical data and extensively tested across DNA structural space.

Machine-learning systems now provide another route. AlphaFold 3 and related models can predict structures involving proteins, DNA and RNA with remarkable speed, but they are learned systems whose capabilities emerge from training on biomolecular structural information.

Flux Genome Physics is testing a different route. For the helical-rise result, the calculation does not consume:

The SantaLucia nearest-neighbor data, for example, are used on the validation side, not as the source of the calculated interaction.

There are, of course, physical and structural inputs: chemical identity, the Watson–Crick pairing face, the canonical B-DNA twist, and geometric closure constraints. The helical rise is not one of them.

That distinction matters. The claim is not that computation can occur without inputs. The claim is that the property being tested is not learned or fitted from the property data it is being compared against.

Why not just use quantum chemistry?

There is already a route from fundamental electronic physics into DNA: quantum chemistry. Density functional theory has been applied to DNA duplexes, including full Dickerson-dodecamer calculations, and modern linear-scaling DFT methods can handle surprisingly large biological systems.

That is important, because the distinction is not “nobody can calculate DNA with physics.” They can. The distinction is computational economics.

Large biological systems challenge conventional electronic-structure methods because every increase in system size increases the cost of resolving the electronic problem. Specialized linear-scaling methods have pushed that boundary dramatically, but published DNA work still describes full-system quantum treatment as requiring significant computational resources.

That makes quantum chemistry extraordinarily valuable for high-detail analysis. It makes it much harder to use as the inner loop of a search engine exploring enormous numbers of sequence states, mutations, oligos, structural alternatives or redesigns.

This is the gap Genome Physics is trying to enter. Not physics or speed, but eventually physics at search speed.

Because if the underlying calculation remains compact enough to accelerate, the end state is fundamentally different from making a single DNA simulation faster. It becomes possible to search physical sequence space.

From simulating DNA to searching it

This is where the Genome Physics program becomes strategically interesting. Consider a mutation. The conventional biological question is often: is this mutation associated with disease, resistance, altered expression or another measured outcome?

Those are important questions. But they sit high in the causal chain. Genome Physics starts lower:

What changed physically?

Did the mutation change pairing, stacking, local geometry, backbone strain, opening tendency, mechanical response, interaction with a protein or ligand, or the energy ordering of competing structural states?

That is the physical delta. Once that layer exists, it becomes an input into higher biological reasoning rather than being replaced by it.

The same logic applies to oligos and guides. Instead of asking a model only which sequence resembles historically successful candidates, a future Genome Physics workflow could enumerate candidates and ask the underlying physical engine which ones form the intended duplex most favorably, which discriminate most strongly against a mismatch, which create self-structure, which remain robust across a resistance mutation, and which local modifications improve the physical state while preserving the required biological constraints.

That was the original reason for designing Genome Physics as a compiler, not a single prediction model. AI can propose candidates. Enumeration can propose candidates. Databases can propose candidates. Human scientists can propose candidates. But the physical property used to promote one candidate over another should still be recomputed from the physical system.

The missing quadrant

Computational life science currently has several extraordinarily powerful tool families. Quantum methods offer deep physical grounding, but their cost limits brute-force exploration. Molecular mechanics and coarse-grained models can explore larger systems and longer times, but require carefully designed parameterizations. Machine learning offers extraordinary inference speed, but it learns from existing examples and inherits the strengths and boundaries of those examples.

These approaches are not enemies. We expect all of them to remain useful. The interesting possibility is that there may be a fourth operating point:

That bottom-left cell is the bet: deterministic, mechanistic, searchable, and increasingly derived from one physical foundation rather than a new fitted model for every new biological property.

A small result with a much larger destination

A helical rise of 3.16 Å is not a genome engine. It is not a therapeutic designer. It is not whole-cell biology. And it is certainly not “biology solved.”

It is more useful than that. It is evidence that the foundational question is worth continuing to ask:

How much biological structure can emerge from the physics itself before we have to teach the computer what biology normally looks like?

When we published the Genome Physics roadmap in July, it was still primarily a scientific architecture and benchmark contract. Its first dependency was the Nucleotide and Duplex Core. Now there is a complete backbone-first B-DNA research pipeline, and one of the defining dimensions of the double helix is beginning to emerge from the physical closure of the molecule itself.

The next destination is not simply better reproduction of known DNA. It is using the same machinery to calculate changes.

That is when Genome Physics becomes more than a new way to simulate DNA. It becomes a new way to compute with it.

Explore the program

Flux Genome Physics is the DNA, RNA, mutation and therapeutic-sequence frontier of the Matter Computing program. The operating principle remains simple: sequence identifies the matter; state determines the physical problem. And the commercial objective remains deliberately bounded: reduce the physical design space before expensive experiments begin.