Levels of Evidence in Protein Structure Prediction: AlphaFold, Confidence, and Experimental Boundaries

1. What information AlphaFold 2 connects

AlphaFold 2 connects sequences, multiple sequence alignments, and residue-pair representations to three-dimensional coordinate prediction, using evolutionary and geometric information through its network architecture and training approach. Prediction does not directly deduce a unique, experimentally established native conformation from a sequence; input databases, templates, versions, and the target state still affect research interpretation. Structural evaluation in the paper and subsequent biological uses must be described separately.[1] Highly accurate protein structure prediction with Alpha

2. Local confidence is not a probability of function

EMBL-EBI interprets pLDDT as per-residue local confidence on a scale from 0 to 100. Values above 90 generally correspond to very high local accuracy; values above 70 generally support a reasonably good backbone prediction, although some side chains may be inaccurate. These interpretations do not mean that “this protein has a 90% probability of performing a particular function,” nor are they experimentally measured error bars. Dividing the score by 100 does not justify arbitrarily treating it as an event probability.[3] pLDDT: Understanding local confidence

3. How to interpret low-confidence segments

Different regions of the same protein chain can have different confidence levels. Low scores may indicate prediction uncertainty or may be associated with flexibility, disorder, or other conditions; a color map alone cannot establish the specific cause. This article recommends examining local regions according to the research question and incorporating independent information. Taking a single average over the entire protein can hide differences between critical and noncritical positions.[3] pLDDT: Understanding local confidence[1] Highly accurate protein structure prediction with Alpha

4. AlphaFold 3 expands the range of entities modeled

AlphaFold 3 uses a diffusion module to generate atomic coordinates directly, addressing a wider range of biomolecular interactions in a unified manner. The original paper also discusses the challenge of generative methods producing seemingly plausible structures in disordered regions, along with the training design used to address it. Broader capabilities do not mean that every generated conformation is correct; different tasks also have different sampling, ranking, and evaluation conditions.[2] Accurate structure prediction of biomolecular interacti

5. Prediction, interaction, and function cannot be linked by successive inferential leaps

A high-confidence structural hypothesis can help formulate experimental questions, but cannot by itself establish binding strength, reaction rates, or cellular effects. Beyond local geometry, relative conformations, environmental conditions, and dynamics also affect interpretation. Figure 1 organizes the evidence workflow linking prediction, confidence information, and independent experiments; Figure 2 is an explicitly constructed two-domain matrix. Each additional category of scientific conclusion requires corresponding independent measurements.作者分析与复核建议 / Author analysis

6. What to retain for a computational structural result

At a minimum, record sequences and modifications, model and database versions, template and alignment sources, random seeds, sampling counts, ranking rules, confidence metrics, and raw outputs. If only the best conformation is shown, explain how it was selected and whether the other candidates agree with one another. Reproduction requires an explicit definition of “the same run conditions”; retaining only an attractive structural rendering is insufficient. These are recommendations for reading the literature and maintaining records.作者分析与复核建议 / Author analysis

7. Defining the conclusions this article can support

This article summarizes model mechanisms and official interpretations of metrics. It makes no judgments about particular proteins, variants, or drugs, and has neither run AlphaFold nor generated new coordinates. The values in Figure 2 in this version are entirely constructed, replacing the confidence color-band illustration used in the previous version; they explain only the distinction between intra-domain and inter-domain information. New scientific claims should be assessed against original measurements, metric definitions, and task conditions, without treating a structural image as final evidence.[3] pLDDT: Understanding local confidence

The scale at which a metric applies is key to interpretation

Local structural metrics and metrics of relative inter-domain positions address different spatial scales. Even if two domains are each individually very stable, their relative placement may remain uncertain. Compressing all these issues into a single average score can cause readers to overlook relationships that actually affect research judgments. Confidence information should be selected according to the question, and raw outputs should be retained, rather than presenting only a visually polished final view.作者分析与复核建议 / Author analysis

What subsequent validation needs to answer

If a research claim concerns a particular interface or functional change, subsequent evidence should directly address that claim. Computational results used to propose hypotheses should be distinguished from new measurements used to test them, and circular justification using the same source should be avoided. Functional interpretations without independent evidence should explicitly retain the status of hypotheses. This article provides only principles for organizing evidence; it gives no specific experimental operating procedures and makes no extrapolations to clinical use.作者分析与复核建议 / Author analysis

8. Confidence and known error are not interchangeable

pLDDT is model-predicted information about local quality, not a measurement obtained by subtracting predicted coordinates from a known ground truth atom by atom. PAE expresses uncertainty in relative positions under specified alignment conditions and can help distinguish cases in which intra-domain structure is relatively reliable but inter-domain orientation is uncertain. High local confidence does not imply that the overall assembly, binding interface, or a particular conformation is correct. When interpreting these quantities, retain the model version, output definitions, and sequence context; scores from different methods or versions cannot be directly treated as equivalent probabilities without a basis for calibration.[1] Highly accurate protein structure prediction with Alpha[3] pLDDT: Understanding local confidence

9. Why alignment changes structural error

Coordinate error depends on which degrees of freedom are removed before comparison. RMSD after global rigid alignment is affected by flexible inter-domain motion; aligning each domain separately can instead conceal errors in relative inter-domain positions. Both comparisons can be meaningful, but they answer different questions. The atom set, missing residues, symmetry handling, and alignment strategy should be specified, and local and global results should both be reported when needed. When multiple experimental conformations exist, a single reference may also be insufficient to represent the true conformational distribution.[1] Highly accurate protein structure prediction with Alpha

10. What inferences still separate structure from function

A static structural hypothesis cannot by itself establish binding strength, reaction rates, localization under cellular conditions, or clinical effects. Complex or ligand predictions can generate testable interfaces, but whether binding occurs, persists under relevant conditions, and changes function each requires its own experiments. Researchers should list in advance which structural features will be tested, which observations will be used to judge them, and which alternative explanations remain. AF3 expands the range of entities modeled; it does not remove requirements concerning data coverage, conformational diversity, or experimental validation.[2] Accurate structure prediction of biomolecular interacti[1] Highly accurate protein structure prediction with Alpha

11. Verification requires retaining inputs and experimental context

The sequence version, homology information, template dates, and chain and ligand definitions on which a prediction depends all affect verification. Comparisons of approaches must guard against unfair advantages arising from target structures or similar information available before or after a cutoff date. Experimental results must also retain records of constructs, conditions, and structure-determination methods; structural differences between different constructs cannot simply be attributed to model error. For low-confidence regions, disorder, flexibility, and modeling failure can all be considered as possible explanations, without directly equating uncertainty with absence of function.[1] Highly accurate protein structure prediction with Alpha[2] Accurate structure prediction of biomolecular interacti[3] pLDDT: Understanding local confidence

Core equations and derivations

Methods and evidence workflow

Analytical figures and numerical examples

Recalculable worked example

The constructed matrix sets within-domain values to 2, between-domain values to 12, and diagonal values to 0, to distinguish local information from relative orientation. It is not a PAE output for any protein, and its colors cannot be used to infer binding, function, or drug efficacy.

Comparison of methods

Verification and experimental plan

1. Retain the exact sequence, construct, chains and ligands, model version, parameters, and template dates; create a new record when input versions change.
2. Examine local confidence and inter-domain uncertainty separately; do not replace analyses of interfaces, flexible regions, and critical sites with a single overall score.
3. If an experimental reference is available, fix the atom set, symmetry handling, and alignment scheme, and report global and relevant local metrics separately.
4. Propose structural hypotheses that can be experimentally falsified, and define measurements, controls, and decision rules for support or lack of support in advance.
5. Retain experimental conditions separately for binding, activity, and cellular phenomena, and record negative results and alternative explanations.
6. For every functional conclusion, identify individually whether it comes from prediction, existing literature, or actual experiments in the current work; this article provides only verification methods.

Scope and limitations

• The illustrative matrix is not a prediction for any real protein.
• Confidence cannot replace structural or functional experiments.
• Conformations, inputs, and experimental conditions affect comparisons.
• No new molecular experiments or judgments about drug efficacy or clinical effects are provided.

Mapping primary sources to claims

1. Highly accurate protein structure prediction with AlphaFoldJumper et al. · 2021-07-15Nature 2021; network, Fig. 2 and limitations
2. Accurate structure prediction of biomolecular interactions with AlphaFold 3Abramson et al. · 2024-05-08Nature 2024; diffusion module and limitations
3. pLDDT: Understanding local confidenceEMBL-EBI Training · Official AlphaFold training; local confidence interpretation