Revolutionizing protein science: AI as an observatory of physical organization in protein space

Artificial Intelligence (AI) is transforming the landscape of biological sciences, especially in protein research. Modern models like AlphaFold have achieved remarkable predictions of protein structures, while protein language models capture subtle evolutionary and functional signals embedded in sequence data.

Protein space is vast, yet the proteins used by living systems are not randomly distributed within it. Proteins that can fold, remain stable, perform biological functions, and survive evolutionary selection occupy only limited regions shaped by physical constraints, evolutionary filtering, and functional requirements. Because protein space is nonuniform, constrained, and statistically learnable, the success of AI models suggests more than improved predictive accuracy. It also indicates that these models may have captured deeper regularities underlying the protein world.

The review defines predicted structures, confidence scores, sequence embeddings, mutation-effect estimates, inverse-design scores, and generative ensembles as “AI-derived observables.” These outputs are not direct experimental measurements, nor can they be simply equated with real conformations, physical energies, or biological mechanisms. However, when compared, calibrated, and validated, they can serve as important readouts for exploring protein space and extracting interpretable physical and biological information from model results.

Within this framework, the review discusses several classes of models that together form an AI-based observatory of protein systems. Classical physical and statistical approaches, including molecular dynamics, energy-based modelling, multiple sequence alignments, and direct coupling analysis, provide reference frames for interpreting model outputs. Structure-prediction models infer three-dimensional organization from sequence and evolutionary information while also offering confidence and uncertainty readouts. Protein language models learn patterns of evolutionary variation, structural compatibility, and functional constraints from large-scale sequence data. Generative and inverse-design models explore which sequences, structures, or assemblies are feasible under learned rules, helping to map the designable regions of protein space.

The review further highlights three large-scale patterns revealed by AI-derived protein data. First, predicted-structure databases are transforming the protein universe into a searchable structural map, enabling the identification of remote structural relationships and fold-level neighborhoods beyond sequence similarity. Second, proteome-scale analyses of predicted structures make it possible to systematically examine how folding topology relates to native-state dynamics, flexibility, stability, and functional specialization. Third, multimodal representations are bringing sequence, structure, and function into shared computational spaces, supporting remote homology detection, functional annotation, enzyme prediction, and cross-modal retrieval, while also prompting deeper questions about how sequence, structure, dynamics, and function are jointly organized through evolution.

At the same time, the authors emphasize that AI-derived signals must be interpreted with caution. Their meaning depends on training data, model architecture, input information, and downstream filtering, and they cannot be directly used as scientific conclusions without calibration. To improve interpretability, the review discusses strategies such as confidence and uncertainty analysis, perturbation analysis and mutation scoring, contrastive scoring across conformational states, representation decomposition, and physically informed probes including MSA subsampling, masking, frustration analysis, and ensemble refinement. These approaches help connect AI outputs with folding, conformational dynamics, evolutionary filtering, functional response, and design feasibility.

Experimental validation remains a central step in AI-driven protein discovery. Benchmarks, deep mutational scanning, structural measurements, binding assays, functional tests, and prospective design experiments are needed to determine which AI-derived patterns reflect real protein behavior and which arise from model bias or limited data coverage. Experiments and simulations do more than validate individual predictions. They form feedback loops through which model-derived signals are corrected, extended, and converted into reliable scientific knowledge.

Overall, this review situates AI within a broader framework of scientific discovery. AI is not merely a prediction engine, but a new observational interface for protein science. Just as systematic observations in physics became transformative only when converted into interpretable regularities and theoretical principles, AI-derived protein data must be calibrated by physical constraints and experimental validation before they can serve as scientific evidence for understanding the organization of protein space. The next stage of AI-driven protein research may therefore depend less on single impressive predictions and more on calibrated, interpretable, and testable maps of protein space.

Citation: Yuxiang Zheng, Zecheng Zhang, Yuxiao Wang, Wenbin Kang, Weitong Ren, Qian-Yuan Tang. From Prediction to Discovery: AI as an Observatory of Physical Organization in Protein Space[J]. AI for Science. DOI: 10.1088/3050-287X/ae78ea