Egret-1: Pretrained Neural Network Potentials for Efficient and Accurate Bioorganic Simulation (2504.20955v5)

Abstract: Accurate simulation of atomic systems has the potential to revolutionize the design of molecules and materials. Unfortunately, exact solutions of the Schr\"odinger equation scale as O(N!) and remain inaccessible for systems with more than a handful of atoms, forcing scientists to accept steep tradeoffs between speed and accuracy and limiting the reliability and utility of the resultant simulations. Recent work in machine learning has demonstrated that neural network potentials (NNPs) can learn efficient approximations to quantum mechanics and resolve this tradeoff, but existing NNPs still suffer from limited accuracy relative to state-of-the-art quantum-chemical methods. Here, we present Egret-1, a family of large pretrained NNPs based on the MACE architecture with general applicability to main-group, organic, and biomolecular chemistry. We find that the Egret-1 models equal or exceed the accuracy of routinely employed quantum-chemical methods on a variety of standard tasks, including torsional scans, conformer ranking, and geometry optimization, while offering multiple-order-of-magnitude speedups relative to legacy methods. We also highlight important lacunae for future NNP research to investigate, and suggest strategies for building future high-quality models with increased scale and generality.

• The paper presents Egret-1, a novel pretrained neural network potential that achieves quantum accuracy with reduced computational expense.
• It demonstrates superior performance on benchmarks like GMTKN55 and ROT34 by lowering errors in energy and geometric predictions.
• Egret-1 employs higher-order equivariant architectures to model complex bioorganic simulations without the need for extensive modifications to datasets.

Egret-1: Neural Network Potentials for Bioorganic Simulations

The paper introduces Egret-1, a family of neural network potentials (NNPs) that are designed to improve the simulated accuracy of quantum mechanical systems at reduced computational costs. Situated at the crossroads of machine learning and computational chemistry, Egret-1 enhances the capability to model complex bioorganic materials. It is noteworthy that this advancement occurs without necessitating sweeping modifications in dataset sizes or underlying model architectures, which showcases the efficiency of the model in optimizing existing resources for an accurate emulation of complex systems.

Egret-1 builds on the higher-order equivariant MACE architecture and offers several versatile pre-trained models. These models are applicable to a broad swath of chemical space, including organic and biomolecular chemistries, and promise quantum mechanics-level performance across various standard tasks like torsional scans and conformer ranking, all while delivering speedups of multiple orders of magnitude over existing methods.

Strong Numerical Results and Comparative Performance

Egret-1's performance is quantitatively assessed against various benchmarks, revealing several areas of marked superiority over other methods. For example, in the GMTKN55 dataset, Egret-1 shows improved weighted total mean absolute deviation (WTMAD-2) scores compared to typical neural network potentials such as MACE-MP-0b2-L and Orb-v3. While DFT methods like B97-3c remain competitive, Egret-1 offers a comparable accuracy at a fraction of their computational expense. Notably, Egret-1 models excel in tasks involving molecular geometries, with bench tests such as the ROT34 proving that Egret-1 achieved superior accuracy when predicting rotational constants, surpassing well-established density functional theory (DFT) methods.

In conformational and torsional profile predictions, crucial for computer-assisted drug design, Egret-1 models showcase reduced mean absolute error (MAE) and root mean square error (RMSE) in datasets such as Folmsbee and TorsionNet206, when juxtaposed with current top-performing NNPs. Their efficacy extends to challenging benchmarks like ROT34 for geometries and Wiggle150 for strained conformers, where the Egret-1 suite shines in accurate energy predictions and reproducing geometric constants respectively.

Methodologies and Implications

Methodologically, Egret-1 models were trained using diverse datasets including the MACE-OFF23 and VectorQM24, which yield some insights into data sensitivity and its effects on model performance. Here, the integration of complex dataset diversity can inadvertently degrade rather than improve overall model accuracy, signaling a need for refined strategies in data aggregation and training regimens. This is evidenced particularly under Hessian matrix sensitivity benchmarks such as VIBFREQ1295.

The theoretical contributions of Egret-1 are also noteworthy. Using neural message-passing mechanisms, Egret-1 demonstrates permutation invariance and SO(3) equivariance, thereby enabling accurate energy and force predictions based on rotational properties of molecules. This architectural feature could broadly affect how chemists simulate directional forces inherent in complex tasks, such as modeling catalytic processes or predicting material properties.

Future Scope and Development

Despite its strengths, Egret-1 and its variants confront limitations in terms only supporting specific elements and neutral, closed-shell molecules. Moreover, the models presently support gas-phase calculations exclusively, hence are less directly applicable within solvated environments – a key aspect for simulations involving biochemical processes.

Looking forward, development of enhanced training strategies, such as dynamic weighting of datasets, pre-training or fine-tuning protocols might bridge these gaps. Such advancements could bolster generalizability across broader chemical domains, essential for propelling Egret-1 from computational landscapes into diverse experimental workflows. Moreover, cross-disciplinary methodologies merging existing density-functional theory with neural network-inferred potentials may afford additional avenues to tap into the latent potential of machine learning in chemistry.

In conclusion, Egret-1 epitomizes a significant stride forward in bioorganic simulations, marrying the computational efficiency of machine learning with domain-specific quantum chemical principles, thereby extending feasible simulation boundaries for researchers in fields stretching from drug discovery to advanced materials development.