---
title: 'Nutmeg: Biomolecular & Bayesian Systems'
url: https://www.emergentmind.com/topics/nutmeg
type: topic
---

# Nutmeg: Biomolecular & Bayesian Systems

In recent arXiv literature, **Nutmeg / NUTMEG** denotes two distinct research systems. In biomolecular machine learning, **Nutmeg** is a family of machine-learning potentials for drug-like small molecules, peptides, and noncovalent complexes, trained on the SPICE v2 quantum-chemistry dataset and based on the TensorNet architecture [2406.13112]. In natural language processing and annotation science, **NUTMEG**—**Nuanced Understanding of annoTation by MultiplE Groups**—is a Bayesian model for separating noisy annotations from systematic subgroup disagreement in crowdsourced labeling workflows [2507.18890]. The shared name therefore spans two unrelated technical lineages: one in atomistic simulation and one in probabilistic modeling of human annotation.

## 1. Disambiguation and scope

The term appears in at least two technically unrelated senses in the cited literature.

| Usage | Domain | Core objective |
|---|---|---|
| Nutmeg | Biomolecular machine learning | Predict potential energies and forces for small molecules, peptides, and noncovalent complexes |
| NUTMEG | NLP / annotation modeling | Remove noisy annotations while preserving systematic disagreements across annotator subpopulations |

The biomolecular **Nutmeg** system is introduced alongside version 2 of the SPICE dataset and is explicitly positioned as a set of equivariant message-passing potentials intended to achieve near–quantum-mechanical accuracy and routine molecular dynamics throughput on small systems [2406.13112]. The annotation-model **NUTMEG** is introduced as a Bayesian aggregation framework that models annotator competence, spam, and subgroup-specific ground truths in order to preserve meaningful disagreement rather than collapsing all deviation into error [2507.18890].

A plausible implication is that the shared name does not denote a common methodology; rather, it is a homonym across two research areas with different data modalities, inductive biases, and evaluation regimes.

## 2. Nutmeg as a biomolecular machine-learning potential

In biomolecular machine learning, the Nutmeg family implements **equivariant message-passing networks** to predict **potential energies and forces** for **drug-like small molecules, peptides, and their noncovalent complexes** [2406.13112]. Its stated goals are twofold: **better than 1 kcal/mol on average** across a broad range of chemistries, including charged and polar species, and **milliseconds per force evaluation on modern GPUs** so as to enable routine molecular dynamics of small systems.

Nutmeg is trained on **SPICE v2**, a dataset described as containing **2.0 million conformations** spanning **17 elements**, **explicit solvent clusters**, **ligand–amino acid pairs**, and both **neutral and charged species** [2406.13112]. The model was designed partly in response to the fact that earlier machine-learning potentials and datasets often omitted **charged molecules**, **non-equilibrium conformations**, or **noncovalent interactions**.

Architecturally, Nutmeg builds on the **TensorNet framework** with three main modules: an **embedding layer**, a stack of **interaction layers**, and an **output layer** [2406.13112]. The interaction layers perform successive rounds of message passing over neighbors within cutoff $r_c$, and a model with $n$ interaction layers has a receptive field of $(n{+}1)\,r_c$. The output layer computes atomic energy contributions $E_i$, which are summed into the total potential energy.

The energy is decomposed as

$$
E_\text{total} \;=\; \sum_{i} E_{i}\bigl(\text{local environment}\bigr) \;+\; E_\text{ZBL},
$$

with the short-range repulsive term

$$
E_\text{ZBL} \;=\; \sum_{i<j} w(r_{ij})\,\mathrm{ZBL}(r_{ij}).
$$

The cutoff function is

$$
w(r) =  \begin{cases} \frac12\bigl[1 + \cos(\pi\,r/r_\text{max})\bigr], & r<r_\text{max},\\ 0, & r\ge r_\text{max}. \end{cases}
$$

This repulsive term is described as a **short-range Ziegler–Biersack–Littmark repulsion** suppressed beyond the sum of covalent radii [2406.13112]. Its role is to prevent **unphysical collapse** in simulation.

## 3. Charge injection, training protocol, and dataset composition

A defining feature of Nutmeg is its **charge-injection mechanism**, introduced to improve accuracy on **charged and polar molecules** [2406.13112]. The procedure is specified as follows: compute **Gasteiger atomic partial charges** once for each molecule, concatenate each atom’s one-hot element encoding with the corresponding charge, and map the augmented feature vector through a learnable linear transform to obtain the initial embedding,

$$
\mathbf{h}_i^{(0)} = W\bigl[\mathrm{one\_hot}(Z_i)\,\|\,q_i\bigr].
$$

No additional Coulomb kernel is imposed; the network is allowed to exploit the injected charges implicitly through message passing. The paper states that no explicit charge regularization term is used because the partial charges are **fixed inputs rather than predicted quantities** [2406.13112].

The model is trained to reproduce both reference energies and forces using the loss

$$
\mathcal{L} =
\bigl\lVert E_\text{pred}-E_\text{ref}\bigr\rVert_{2}^{2}
+\lambda\,
\bigl\lVert \mathbf{F}_\text{pred}-\mathbf{F}_\text{ref}\bigr\rVert_{2}^{2},
$$

with $\lambda = 0.01$ in units of $(\mathrm{kJ/mol})^{-2}$ [2406.13112].

SPICE v2 is described as comprising several sources of geometries, including **PubChem drug-like molecules** with approximately **28 k molecules** and **1.4 M conformers**, **ligand–amino acid pairs** with approximately **10 k ligands** and **194 k conformers**, **solvated PubChem clusters** with approximately **14 k solutes** and **63 k conformers**, **water clusters** with **30 molecules** and **1 k conformers**, as well as **dipeptides, solvated amino acids, DES370K monomers/dimers, and reran DFT outliers** [2406.13112]. All geometries were computed at the **wB97M-D3(BJ)/def2-TZVPPD** level with **Psi4**, and forces were obtained analytically.

The training split holds out a random **5% of conformations (100 k)** for validation, leaving **1.9 M** conformations for training [2406.13112]. Shared hyperparameters across the three reported model sizes are **AdamW**, **batch size 64**, initial learning rate $\alpha_0=5\times10^{-4}$ reduced by **0.8 on plateau** to a minimum of **$10^{-5}$**, **100 total epochs**, **cutoff $r_c=1.0$ nm**, and **ZBL cutoff $r_\text{max}$ equal to the sum of Slater covalent radii**. The three variants—**Nutmeg-small**, **Nutmeg-medium**, and **Nutmeg-large**—differ in embedding/interact widths and number of interaction layers.

## 4. Empirical behavior of the biomolecular Nutmeg models

On the held-out SPICE v2 validation set, the reported mean absolute errors in total energy are **2.60 kJ/mol** for Nutmeg-small, **2.19 kJ/mol** for Nutmeg-medium, and **1.85 kJ/mol** for Nutmeg-large, all below **4.184 kJ/mol (1 kcal/mol)** [2406.13112]. The error grows modestly with system size up to approximately **70 atoms** and then increases sharply, which the paper attributes to the finite message-passing cutoff. Reducing the cutoff from **1.0 nm** to **0.6 nm** moves this inflection to approximately **50 atoms**, reinforcing the dependence on receptive field size.

A separate test set of **7976 conformations** was assembled from molecules not present in SPICE v2: **small ligands (40–50 atoms)**, **large ligands (70–80 atoms)**, **pentapeptides (68–110 atoms)**, and **novel dimers (34–72 atoms)** [2406.13112]. Absolute-energy MAEs on this test set range from **6–15 kJ/mol**, decreasing with model size, while conformational energy differences are substantially more accurate. For example, **Nutmeg-large** achieves approximately **4 kJ/mol MAE on pentapeptides** for conformational $\Delta E$, despite their size.

The effect of charge injection is reported explicitly. Without injected charges, absolute-energy errors grow dramatically with the number of formal charges; with **precomputed partial charges**, this dependence is said to collapse entirely for conformational $\Delta E$ [2406.13112]. A control model using only **formal charges** instead of **Gasteiger partial charges** yields intermediate performance, which the paper interprets as evidence that **detailed partial charges** are the more effective baseline for charged systems.

Simulation stability and speed are also evaluated. Over **1800 short MD runs** of **10 ps** with **1 fs timestep** on **600 test molecules**, the models exhibited **no integrator failures or runaway temperatures**, with only a handful of bond breaks, limited to **hypervalent sulfur motifs**, which were also unstable at the reference DFT level [2406.13112]. On an **NVIDIA RTX 4080**, **Nutmeg-small** reaches approximately **150 ns/day** on a **50-atom molecule**, characterized as approximately a **$10^6$-fold speedup over DFT** but still approximately **$10^3$-fold slower than classical force fields on large systems**.

For torsional energetics, **Nutmeg-large** reproduces **eight representative torsion scans** from the **TorsionNet 500 benchmark** with **barrier-height MAEs of $\lesssim 1.1$ kJ/mol** [2406.13112]. However, in a **2.2 nm periodic box of water**, **Nutmeg-small** shows poor liquid structure relative to **TIP4P-FB**, **AMOEBA**, and **SWM4-NDP**, indicating that the available **30-molecule water clusters** in SPICE v2 are insufficient for accurate bulk-water modeling.

## 5. Limitations and projected extensions of the biomolecular model

The biomolecular Nutmeg models are explicitly described as **not designed to model bond breaking/forming** [2406.13112]. The use of **fixed partial charges** prevents on-the-fly reparameterization for reactions, and the finite cutoff limits accuracy beyond approximately **70 atoms** unless the architecture is extended.

The paper further states that **bulk-phase properties remain a challenge**, plausibly because larger solvent aggregates or explicit long-range treatments are required [2406.13112]. Proposed future directions include **global or trainable charge-equilibration schemes for dynamic polarization**, extending **SPICE** with **larger solvent aggregates and condensed-phase simulations**, integrating **CUDA Graphs** or **PyTorch 2 compiler optimizations** to reduce GPU overhead, and exploring **higher-order message-passing** or **multi-scale electrostatic kernels** to improve transferability to **macromolecular and periodic systems**.

These statements delimit the intended use case: routine simulation of **small systems** with high fidelity for **charged**, **polar**, and **noncovalent** chemistries, but not reactive chemistry or well-resolved condensed-phase water behavior [2406.13112].

## 6. NUTMEG as a Bayesian model of annotator disagreement

In a separate line of work, **NUTMEG** denotes **Nuanced Understanding of annoTation by MultiplE Groups**, a Bayesian model for crowdsourced annotation in which disagreement is decomposed into **“spam” or noise** and **“signal”**, meaning genuine, systematic disagreement driven by annotator background [2507.18890]. The model is motivated by the observation that conventional aggregation procedures such as **majority vote**, **Dawid & Skene**, **MACE**, and **GLAD** typically treat all deviations from consensus as error, while “learning from disagreement” approaches may over-weight raw noisy labels.

The model assumes **$N$ items**, **$M$ annotators**, **$P$ known subpopulations**, and a label space of size **$L$** [2507.18890]. Each annotator $j$ belongs to exactly one subpopulation $g_j=k$. The latent variables are:

- $T_{i,k}$: the true label for item $i$ according to subpopulation $k$
- $\theta_j$: annotator competence
- $\phi_j$: annotator spam distribution over labels
- $S_{i,j}$: whether annotator $j$ is spamming on item $i$
- $A_{i,j}$: the observed annotation

The priors are given as

$$
T_{i,k} \sim \mathrm{Uniform}(\{1,\dots,L\}),
$$

$$
\theta_j \sim \mathrm{Beta}(\alpha_0,\alpha_0),
$$

with $\alpha_0 = 0.5$ to encourage extremes, and

$$
\phi_j \sim \mathrm{Dirichlet}(\beta_0,\dots,\beta_0),
$$

with symmetric $\beta_0 = 0.5$ [2507.18890].

For each annotation, the spam indicator is sampled as

$$
S_{i,j} \mid \theta_j \sim \mathrm{Bernoulli}(1-\theta_j),
$$

so that $P(S_{i,j}=0)=\theta_j$ corresponds to the annotator following the subgroup truth and $P(S_{i,j}=1)=1-\theta_j$ corresponds to spam. The observation model is deterministic when $S_{i,j}=0$, in which case $A_{i,j}=T_{i,g_j}$, and categorical from $\phi_j$ when $S_{i,j}=1$ [2507.18890]. The joint distribution is written as

$$
P(A,T,S,\theta,\phi) =
\prod_{i=1}^N \prod_{k=1}^P P(T_{i,k})
\prod_{j=1}^M P(\theta_j)P(\phi_j)
\prod_{i=1}^N \prod_{j=1}^M P(S_{i,j}\mid \theta_j) P(A_{i,j}\mid S_{i,j},T_{i,g_j},\phi_j).
$$

The inference procedure uses **Variational Bayes**, with a mean-field variational posterior over $T$, $S$, $\theta$, and $\phi$, optimized via the **Evidence Lower Bound**,

$$
\mathcal{L}(q) = E_q[\ln P(A,T,S,\theta,\phi)] - E_q[\ln q(T,S,\theta,\phi)].
$$

The update equations are given in the paper in the familiar variational form $q \propto \exp(E[\ln \text{joint}])$, with conjugate updates for $q(\theta_j)$ as a **Beta** distribution and for $q(\phi_j)$ as a **Dirichlet** distribution [2507.18890]. Because each annotator is tied to a known subpopulation, all expectations involving truth use the subpopulation-specific variable $T_{i,g_j}$.

## 7. Experimental findings and downstream consequences of the annotation model

The NUTMEG annotation model is evaluated in three settings [2507.18890]. In **synthetic data (Exp 1)**, the setup uses **$M=150$ annotators**, split **80% majority** and **20% minority**, over **$N=500$ binary items**. A proportion $\delta$ of items is made **divisive**, meaning $T_{i,\mathrm{maj}} \neq T_{i,\mathrm{min}}$, and each item receives **5 random annotators**. Baselines are **Majority Vote**, **Dawid & Skene (1979)**, **MACE (Hovy et al. 2013)**, **LFC (Raykar et al. 2010)**, and **BCC (Kim & Ghahramani 2012)**. As $\delta$ increases, the traditional methods remain accurate on the majority label at approximately **95%** but collapse on the minority to approximately **50% at $\delta=1$**. By contrast, NUTMEG remains high for both groups even when $\delta=1$, with **minority accuracy approximately 90–95% at low spam** and approximately **80–85% at spam = 0.25**. Its competence estimates correlate at approximately **$r=0.81$** with true competence, compared with **0.58 for MACE**.

In **subpopulation size analysis (Exp 2)**, with **global spam = 0.1** and **divisiveness $\delta = 0.2$**, the minority proportion $\alpha$ varies over **$\{0.1,\dots,0.5\}$** and annotations per item $m$ over **$\{3,\dots,15\}$** [2507.18890]. The reported result is that maintaining minority accuracy at approximately **92%** requires about **$m \approx 5$** when **$\alpha = 0.3$**, but **$m > 15$** when **$\alpha = 0.1$**. The paper interprets this as evidence that practitioners must **oversample small subpopulations** if accurate minority truths are desired.

In **downstream modeling (Exp 3)**, the real dataset is **POPQUORN (Pei & Jurgens 2023)** with annotations for **politeness** and **offensiveness** on **1–5 Likert** scales, binarized at **$\ge 3$**, together with annotator demographics including **age, race, gender, education** [2507.18890]. The base encoder is **ModernBERT**. Two settings are evaluated: aggregated labels produced by **majority vote**, **MACE**, or **NUTMEG’s per-subpopulation truths**, and a disaggregated multi-task model with one head per annotator. The metric is **Jensen–Shannon divergence (JSD)** between the predicted label distribution for each subpopulation and the empirical distribution of disaggregated test labels in that subpopulation.

For **politeness**, the reported mean JSD is approximately **0.04** for **NUTMEG-aggregated** labels, approximately **0.06** for **MACE**, approximately **0.07** for **majority vote**, and approximately **0.10** for **full disaggregation**, corresponding to a **30–40% reduction in JSD over MACE** [2507.18890]. For **offensiveness**, all aggregation methods improve on raw data, and **NUTMEG still matches or slightly betters MACE**, with **JSD approximately 0.055 versus 0.056**. The paper notes that demographic subgroup signals were weak in the offensiveness setting, so **noise reduction** was the principal gain there.

Taken together, these results support a precise interpretation of NUTMEG in annotation science: it is a **Bayesian wrapper around crowdsourced labels** that infers annotator competence versus spam, recovers one **true label per subpopulation per item**, and does so through **variational Bayes** in a way that preserves subgroup disagreement while reducing noisy annotations [2507.18890]. A plausible implication is that the model occupies an intermediate position between pure consensus aggregation and unrestricted learning from raw disagreement.

Source: https://www.emergentmind.com/topics/nutmeg