khoji parkAI · United Kingdom · 2020
The Protein Puzzle Met a Predictor
The sequence was readable; the shape was harder.
Read in slides1. A folded problem
Proteins are chains of amino acids whose three-dimensional structures influence their functions. Predicting those structures had challenged researchers for decades. A protein's sequence and its folded shape are different descriptions. Predicting the second from the first was a difficult scientific problem.
The sequence was readable; the shape was harder.
2. The question hidden in a sequence
A list of a protein's building blocks is not the same thing as a description of the shape those blocks form. That difference made the prediction task scientifically important. The system needed to connect one kind of information with another, where a visually plausible answer could still require comparison with independent experimental evidence.
The folded object was not written out in plain text.
3. A blind assessment
CASP evaluates predictions against structures determined experimentally but not yet publicly released. That setup tests results against answers participants cannot simply look up. A benchmark allowed predictions to be compared with experimental structures. Independent comparison mattered more than a confident-looking visualization.
The examination had carefully hidden its answer sheet.
4. The new system
DeepMind entered a new AlphaFold system in CASP14 in 2020. Its predictions achieved a major improvement in structure-prediction accuracy. The result reflected a system built for a specific scientific task. Its success could be evaluated against evidence outside the model itself.
The folding problem had acquired a powerful new tool.
5. The result
The assessment’s organisers recognised the advance as a major breakthrough. High accuracy on the benchmark opened useful possibilities for biological research. Predicted structures can help researchers ask further questions. They do not automatically explain every interaction or replace every laboratory experiment.
A prediction could now guide a much better next experiment.
6. A prediction joins the laboratory conversation
Researchers could use a predicted structure to investigate further questions, but its usefulness depended on knowing what the result did and did not establish. A model can help direct work without explaining every interaction of a molecule. The advance expanded what could be investigated rather than ending the need to investigate proteins at all.
The answer had created more experiments to do.
7. The practical limit
Predicting a protein structure does not automatically explain every function, movement or disease. Researchers still need experiments and careful interpretation. The work changed what researchers could investigate at scale. Its value depends on using predictions with appropriate attention to confidence and context.
The new map did not abolish the need to explore.