TL;DR
This study focuses on two multidomain proteins, LicT from Bacillus subtilis and P1 from the Rice Yellow Mottle Virus, that have long resisted structural characterization.
It is found that, while AF3 can perform well in favourable settings, this performance is uneven across applications and its predictions and use of confidence metrics will depend strongly on the specific application area and must be interpreted with respect to training-set overlap.
This mini review traces the evolution of AI-driven methods in protein research, from early residue-contact prediction using coevolutionary information to transformative breakthroughs, the rise of protein language models (PLMs), and the emerging era of generative design and functional modeling.
Predicting protein stability, like changes in melting temperature (ΔTm) caused by mutations, is a critical task in therapeutic protein engineering and drug discovery. This is reflected by a growing solution space, including both AI-based sequence and structure based methods. This paper demonstrates that accurate ΔTm prediction does not require structural input features, but can achieve state-of-the-art results with a careful training design for large sequence-based protein language models. We combine an autoresearch-inspired setup search with controlled ablation studies and show that a well-tuned sequence-only ESM2-650M model [6] outperforms structure-informed methods in our benchmark, achieving the lowest error (MAE/RMSE) and competitive Pearson correlation without pH or structural inputs. We further show that choices such as loss function, pooling strategy, auxiliary supervision, and finetuning regime materially affect performance.
A variant of the U1A protein containing four substitutions to ionizable residues was generated serendipitously due to a miscommunication. Biophysical measurements reveal this variant has twice the helical structure of wild-type U1A and is trimeric, unlike the monomeric wild type. In sharp contrast, structures predicted by deep-learning (AlphaFold2, RoseTTAFold2) and transformer-based tools (OmegaFold, ESMFold) are nearly identical to the wild-type (backbone RMSD < 1 Å). Surprisingly, these models predict ionizable residues buried within the nonpolar core, contradicting established physico-chemical principles. To explore this effect further, we generated sequences containing up to all twelve residues that make up the nonpolar core of U1A. Across thousands of sequences, and depending on the AI model used, the majority of predicted structures contained fully buried ionizable residues while still maintaining the overall U1A fold. We then examined two additional proteins of comparable size, acylphosphatase and the de novo designed TOP7 fold, and observed the same phenomenon: AI models frequently predicted structures with buried ionizable residues that nevertheless retained the parent fold. However, short (50 ns) molecular dynamics simulations with physics-based force fields (CHARMM/AMBER) rapidly relaxed these structures, exposing the ionizable residues. We conclude that while AI-based tools perform exceptionally on natural sequences, they do not reliably encode the physico-chemical principles governing ionizable residue placement. We propose including brief molecular dynamics simulations as a vital validation step for AI-generated structures.
Addressing and predicting ligand-binding sites in protein structures, as well as the prediction of reliable structures of proteins interacting with other proteins, will be pivotal for fully details of structural mechanisms and dynamics.
Modelling sequences both “dry” and in the presence of explicit potassium cations are suggested as a simple, practical way to sample alternative conformations and to expose disordered regions that current predictors tend to over-fold.