Predict a protein structure
Strong transformation of a central digital task; predictions still need confidence assessment and experimental interpretation.
SCIENCE UNDER AUTOMATION
AI in Structural biology: capabilities and evidence
82/84 top-1 domains have correct folds
What has AI demonstrated?
82/84 top-1 domains have correct folds
- Test
- MULTICOM CASP16; 58 targets, 84 domain units. TM-score >0.5; 62/84 exceed 0.9.
- Comparison
- Experimental structures; AF3-server is also compared.
- System and version
- MULTICOM / CASP16 · 2025
- Evidence checked
- 2026-10-20
84 domain units from 58 targets are not independent proteins. Additional sampling and ranking differs from stock AlphaFold3; no full-field claim.
What does the field require?
Predict a protein structure from sequence
- Prepare sequence and alignment inputs
MULTICOM includes engineered alignment and sampling choices.
MULTICOM / CASP16 - Predict and rank a fold
Blind CASP16 demonstrates strong tertiary structure prediction.
MULTICOM / CASP16 - Determine whether the prediction is adequate
Structural similarity and confidence do not establish experimental validity.
MULTICOM / CASP16
Establish a functional structural mechanism
- Resolve conformations and interactions
Ensemble work uses experimental constraints.
Experiment-guided AlphaFold3 - Obtain discriminating measurements
Autonomous experimental acquisition is not tested in these sources.
Experiment-guided AlphaFold3 - Validate a mechanistic explanation
Whole mechanistic research is not evaluated.
Experiment-guided AlphaFold3
Research tasks
Can AI replace the whole workflow?
The results above measure specific tasks. The field assessment also requires completion of the connected workflow, including the requirements listed here. A field-wide replacement percentage has not been established.
Read the assessment method