JuliaCon 2026

Protein structures are determined experimentally by X‑ray crystallography, electron microscopy, or NMR, or they are predicted with AlphaFold or RoseTTAFold [1,2]. Either way, the resulting protein models often contain missing atoms, distorted bond lengths, or unrealistic side chain orientations generating severe steric clashes and steep energy gradients. Energy minimization is thus essential to achieve physically realistic conformations. While classical minimizers (e.g., conjugate gradient, quasi-Newton) are standard, we demonstrate how machine learning (ML) optimizers can be effectively adapted to this domain.
We introduce a novel mini-batching strategy for ML-based minimization (e.g., SGD, ADAM), implemented in BiochemicalAlgorithms.jl, our library for molecular analysis and simulation [3]. The library provides molecular mechanic functionalities to evaluate energy gradients and corresponding forces, which are essential for energy minimization. Our mini-batching strategy partitions the force‑field contributions by atom‑pair groups, allowing the energy and gradient calculations to be performed on small batches corresponding to random selected portions of the molecule.
Furthermore, a major contribution of this work is a systematic assessment of different optimization algorithms by leveraging the unified Optimization.jl ecosystem—an interface that connects to more than 25 optimization libraries [4].
Our approach supports the possibility of interchanging solvers including SGD and ADAM or alternative ML solvers for energy minimization of proteins without custom code per algorithm.
Using a representative set of proteins from each of the five top‑level SCOP classes, we compared runtime and convergence precision across the solvers. The results highlight which algorithms deliver the best trade‑off between speed and final energy for different structural families.
Overall, the talk will demonstrate how the proposed mini‑batching scheme opens a path towards ML‑augmented protein‑structure refinement and how a single, extensible optimization interface can streamline the evaluation of diverse minimization strategies for protein structure relaxation.

[1] J. Jumper et al., “Highly accurate protein structure prediction with AlphaFold,” Nature, vol. 596, pp. 583‑589, 2021, doi: 10.1038/s41586‑021‑03819‑2.
[2] M. Baek et al., “Accurate prediction of protein structures and interactions using a three‑track neural network,” Science, vol. 373, pp. 871‑876, 2021, doi: 10.1126/science.abj8754.
[3] J. Leclaire et al., “Structure‑based bioinformatics with BiochemicalAlgorithms.jl,” in Proceedings of the JuliaCon Conferences, vol. 7, no. 78, p. 188, 2025, doi: 10.21105/jcon.00188.
[4] V. K. Dixit and C. Rackauckas, “Optimization.jl: A unified optimization package (version v3.12.1),” Zenodo, Mar. 2024. [Online]. Available: https://doi.org/10.5281/zenodo.7738525.
[5] N. K. Fox, S. E. Brenner, and J. M. Chandonia, “SCOPe: Structural classification of proteins—extended, integrating SCOP and ASTRAL data and classification of new structures,” Nucleic Acids Res., vol. 42, no. D1, pp. D304‑D309, 2014, doi: 10.1093/nar/gkt1240.