Does the Nucleobase Ranking Survive the Choice of Dispersion Model?

Computer simulations of DNA-sensing surfaces routinely conclude that a two-dimensional material can tell the four DNA bases apart. These conclusions usually rest on one simulation recipe. Genomics learned a hard lesson about such single-recipe results. When the same sequencing reads are run through dozens of analysis pipelines, the choice of software explains most of the disagreement in the final answer. This proposal applies the genomics method for measuring that disagreement to quantum-chemistry predictions of biosensors. We treat each combination of density functional theory (DFT, the standard quantum method for computing how electrons bind atoms), dispersion correction (an add-on term for the weak van der Waals attraction that plain DFT misses), damping scheme, and geometry protocol as a "pipeline". We then ask how much of the predicted signal difference between bases is physics and how much is recipe.

The core claim is that a sensing-selectivity prediction should count as established only if the gap between neighboring bases exceeds the spread across dispersion-model pipelines, and that this can be measured the way variant-calling reproducibility is measured. A second, smaller step asks whether the part of the signal that survives across methods is large enough to be read out, given the noise of real amplifiers at room temperature and in cold operation. Cryogenic front-end electronics therefore enter as a noise-budget constraint, not as a core mechanism.

Motivation and Prior Work

The motivating result is A Peculiar Binding Characterization of DNA (RNA) Nucleobases at MoOS-Based Janus Biosensor. It computes adsorption of A, G, C, T and U on the two faces of a Janus molybdenum oxy-sulfide monolayer using PBE plus Grimme's D3-BJ correction. It reports binding energies of −0.61 to −0.91 eV, charge transfers of 0.06–0.19 electrons, and a consistent ranking G > A > C > T > U. The binding is almost entirely van der Waals, so the answer depends most on exactly the term that is most model-dependent. Only one dispersion model was used.

Two papers show why this matters and constrain the design. Anisotropy of thermal transport in phosphorene swaps 12 functionals on one monolayer and finds zigzag thermal conductivity ranging from 15 to 152 W/m/K. That is a one-factor-at-a-time sensitivity study; it does not separate which modeling choice drives the spread, or whether choices interact. Calibration of the Many-Body Dispersion Range-Separation Parameter calibrates many-body dispersion (MBD, which models atoms as coupled oscillators rather than summing pairs). It reaches PBE+MBD mean absolute error of 0.37 kcal/mol on the S66x8 set of 528 gas-phase dimer energies against coupled-cluster references. That paper supplies the strongest single-method competitor. It also constrains the proposal: MBD accuracy was shown for small organic dimers, not for molecules lying on polar, partly metallic surfaces.

The method we borrow comes from genomics. Assessing reproducibility of inherited variants detected with short-read whole genome sequencing crosses 56 aligner–caller combinations with platforms and labs and attributes more than 60% of variance to pipeline choice. Toward more accurate variant calling for "personal genomes" reports concordance across pipelines: 57.4% for SNVs and 26.8% for indels. It also shows that calls shared across pipelines validate far better than calls unique to one pipeline (99.1% versus 60.2% for SOAP-unique SNVs). We reuse three things from this work: the crossed factorial design, the variance decomposition, and the observation that consensus-shared predictions are more trustworthy. BinChill adds a cautionary note. Ensemble reconciliation helps (333 versus 31 high-quality bins against DAS-Tool on CAMI2 oral), but it needs tuned thresholds and outside constraints. We therefore test consensus against a reference rather than assuming it is correct.

High-throughput DFT, as in Two-dimensional materials from high-throughput computational exfoliation, already runs automated workflows with two nonlocal functionals, vdW-DF2-C09 and rVV10. It uses the disagreement between vdW and non-vdW geometries as a descriptor of layered materials. We invert that idea. Here the disagreement is the quantity to be explained, not a tool for flagging structures. No paper in the file joins all three directions, and none quantifies dispersion-model uncertainty for nucleobase adsorption.

Research Program

The program has three parts.

The factorial design. We build a crossed design over five factors:

  1. Base functional: PBE, revPBE, optB88-level GGAs, and a meta-GGA such as TPSS.
  2. Dispersion treatment: none, D3-zero, D3-BJ, MBD at calibrated and default range separation, vdW-DF2-C09, and rVV10.
  3. Geometry protocol: dispersion included during relaxation versus added only after relaxation.
  4. Starting orientation of the base: flat versus tilted, with multiple starts.
  5. Code: a plane-wave code and a real-space code, where the dispersion model allows.

Nonlocal functionals are not separable into "base plus add-on", so they enter as their own levels and the design is partly nested. This is handled with a mixed-effects model, not a full ANOVA.

The systems. Each pipeline computes adsorption energy, equilibrium height, charge transfer and band-gap shift for the five nucleobases on:

  1. both faces of Janus MoOS,
  2. MoS2, anchored by the D3-BJ geometry validation in Comprehensive Phase Diagrams of MoS2 Edge Sites, where D3-BJ cut interlayer error from 10% to 1.5%,
  3. phosphorene, where functional sensitivity is already documented.

The reference and the decomposition. For a subset of flat-lying configurations we build finite cluster models of each surface. We compute coupled-cluster-quality interaction energies on them, in the spirit of the S66x8 references, and treat these as an approximate ground truth. We then decompose the variance of each observable into shares from each factor and their interactions.

The central quantity for a pair of bases $i,j$ is a robustness margin:

$$m_{ij} = \frac{|\bar{E}_i - \bar{E}_j|}{\sigma_{ij}}$$

Here $\bar{E}_i$ is the adsorption energy of base $i$ averaged over pipelines, and $\sigma_{ij}$ is the standard deviation across pipelines of the difference $E_i - E_j$. A margin below one means the selectivity claim is inside method noise.

The readout budget. The last step converts the method-robust charge-transfer difference into a signal on a sensing transistor. For a sensor area with $N$ adsorbed molecules, the charge difference between two bases is roughly $N\,\Delta q$. We compare it with the amplifier noise figures in the file: about 1100 electrons of equivalent noise charge at room temperature versus about 400 electrons for CMOS front ends in liquid argon, from ProtoDUNE-SP. This gives the minimum number of molecules needed per readout under each noise floor. We present this as an order-of-magnitude feasibility bound, not a device design. Aqueous samples cannot sit at 87 K, so the relevant case is cold electronics near a warm sensor. The practical question is whether the robust signal is so small that only cold front ends could resolve it.

Baselines

Naive single-recipe reference. The naive reference is the published single-recipe result, PBE+D3-BJ, reproduced on our systems. It stands for current practice. "Beating" it means showing that its ranking or energy gaps are, or are not, reproduced by the reference calculations and by the pipeline ensemble. It is run on the same structures, supercells and convergence settings as every other pipeline.

Strongest competitor: calibrated PBE+MBD alone. PBE+MBD reports MAE = 0.37 kcal/mol on S66x8 in Calibration of the Many-Body Dispersion Range-Separation Parameter. It is the most accurate single method in the file. If it matches the cluster references on surfaces, the ensemble adds little beyond an error bar. We give it the same geometry-relaxation budget and multiple starts as the ensemble. The ensemble "wins" only if it does one of two things. Either its median predicts reference base-pair gaps with lower error than MBD alone, or its spread identifies which pairs MBD gets wrong. We check the second by testing whether pairs with low margin are the pairs where MBD ranks the bases incorrectly against the reference. For no-dispersion PBE, and for every surface-adsorption quantity here, no reported result was found on Lacuna.

Ablations. Two ablations isolate the key ingredient. The first removes dispersion entirely, confirming that it controls the physics. The second replaces the crossed design with one-factor-at-a-time swaps, as in the phosphorene study. If one-at-a-time sensitivity gives the same variance ranking and margins, the factorial machinery is unnecessary.

Why not just…?

Why not just use the most accurate calibrated method, MBD? This is the real threat. The S66x8 calibration covers gas-phase organic dimers. Polarizable surfaces with charge transfer are exactly where oscillator models and range-separation choices can fail, but the file gives no surface evidence either way. The cluster-reference comparison settles it. If MBD alone matches the references within the ensemble spread, we will report that a single calibrated method suffices and the ensemble is only a diagnostic.

Why not just swap a few functionals, as the phosphorene study did? That tells you that results vary, but not why. It also misses interactions, such as a dispersion model that behaves well only with dispersion-consistent relaxation. The one-at-a-time ablation tests directly whether the extra structure changes any conclusion.

Why not just use force fields or machine-learned potentials? Machine Learning Interatomic Potentials and Long-Range Physics documents that local potentials miss long-range dispersion and polarization without explicit add-ons. That is the very quantity being measured here. A potential trained on one DFT recipe would also inherit that recipe's bias rather than exposing it.

Why not just measure selectivity experimentally? In the end one should. But device measurements mix binding with solvent, coverage and electronics noise. This study states in advance which base pairs any experiment could hope to separate, and how many molecules per readout it would need.

Proposed Experiments and Decisive Checks

The first decisive check is the fraction of pipelines that reproduce the full G > A > C > T > U ranking on each MoOS face. Alongside it we compute Kendall rank correlation between pipelines, which plays the role of caller concordance. A second check tests whether consensus rankings agree with cluster references better than unique-to-one-method rankings. This mirrors the shared-versus-unique validation gap in variant calling.

The variance decomposition then reports what share of adsorption-energy variance comes from each factor: the dispersion model, the base functional, the geometry protocol, and orientation sampling. We then compute the margin $m_{ij}$ for all ten base pairs, per surface.

The readout step converts robust pairs into required molecule counts at the two noise floors. The outcome that would overturn the published picture is that close pairs such as C/T or T/U have margins below one. That would mean current single-recipe claims of full four-base discrimination on these surfaces are not supported. The outcome that would vindicate it is margins well above one across all surfaces, together with MBD–reference agreement. The proposal is designed to report either result.