Folding the unfoldable 2: using AlphaFold and ESMFold to explore spurious proteins

Area of Science:

• Computational biology
• Structural bioinformatics
• Protein structure prediction

Background:

• Spurious protein sequences from gene prediction errors are theoretically unlikely to form stable structures.
• Previous studies indicated AlphaFold2 inaccurately assigned high confidence scores (pLDDT) to short spurious sequences, hindering distinction from real proteins.
• The ability of advanced structure prediction tools to differentiate between genuine and artifactual protein sequences remains a critical question.

Purpose of the Study:

• To evaluate if newer protein structure prediction methods, ESMFold and AlphaFold3, can better distinguish between spurious and genuine protein sequences compared to AlphaFold2.
• To assess the performance of these methods in assigning confidence scores (pLDDT) to short spurious sequences.
• To explore the potential of structure prediction scores for identifying spurious proteins at scale.

Main Methods:

• Comparative analysis of protein structure prediction by ESMFold, AlphaFold2, and AlphaFold3 on known spurious sequences (AntiFam).
• Evaluation of prediction confidence scores (pLDDT and pTM) for sequences of varying lengths.
• Development and application of a Gaussian Process Model using structure prediction scores to identify potential spurious proteins in large datasets (AlphaFold DB).

Main Results:

• ESMFold, AlphaFold2, and AlphaFold3 all assigned high pLDDT scores to short spurious sequences.
• Improved discrimination between spurious and real proteins was observed for sequences exceeding 100 amino acids.
• Analysis of disparate pLDDT and pTM scores led to the identification of potentially novel spurious ORFs and a possibly non-spurious AntiFam entry.
• A Gaussian Process Model utilizing structure prediction scores showed promise in identifying potential spurious proteins at scale when combined with other methods.

Conclusions:

• Current advanced protein structure prediction tools, including AlphaFold3 and ESMFold, continue to assign high confidence to short spurious sequences.
• Sequence length is a critical factor, with better discrimination achieved for proteins over 100 amino acids.
• Structure prediction scores offer a valuable, albeit complementary, approach for identifying spurious proteins, especially when integrated into larger identification pipelines.