Search papers, labs, and topics across Lattice.
This study evaluates AlphaFold 3's ability to predict aptamer structures and aptamer-target interactions by comparing its predictions to experimentally resolved aptamer structures in and outside the Protein Data Bank (PDB). AlphaFold 3 accurately modeled PDB-resolved aptamer structures, including those with noncanonical secondary structure elements, but showed lower confidence and accuracy for non-PDB aptamers, though it could still predict G-quadruplex conformations and localize binding interfaces in some cases. The findings suggest AlphaFold 3 holds promise for aptamer modeling but is limited by biases in its training data due to the scarcity of non-genomic nucleic acid structures.
AlphaFold 3 can now predict aptamer structures, but its accuracy hinges on the availability of similar structures in its training data, revealing a crucial limitation for novel aptamer design.
Aptamers, single-stranded nucleic acids that fold into unique three-dimensional shapes, bind selectively to non-nucleotide target molecules, making them promising ligands for diagnostic and therapeutic applications. The ability to accurately predict folded aptamer structures and their molecular interactions would significantly enhance the rational design of nucleic acid-based affinity reagents. However, predicting the 3D structures of aptamers remains challenging due to their complex folding patterns and limited experimental structure data compared to proteins. AlphaFold 3 is the latest structure prediction tool by Google DeepMind that has recently expanded to include nucleic acids and small molecule targets, offering new possibilities for the direct 3D modeling of aptamer sequences. This study evaluates the accuracy of AlphaFold 3 by comparing its predictions to experimentally resolved aptamer structures in the Protein Data Bank (PDB) and to well-characterized aptamers not included in the PDB. AlphaFold 3 effectively modeled a range of PDB-resolved aptamer structures, including those with noncanonical secondary structure elements such as G-quadruplexes and pseudoknots. For non-PDB aptamers, AlphaFold predictions were considerably less confident yet showed reasonable overlap with experimental data, accurately predicting G-quadruplex conformations and, in some cases, correctly localizing known binding interfaces in aptamer–protein complexes. Despite these attributes, AlphaFold 3 predictions appear limited by biases in its training data, reflecting the relative scarcity and redundancy of nongenomic nucleic acid structures in the PDB. These findings highlight the potential of AlphaFold 3 for aptamer modeling but underscore the need for further refinement to reliably predict complex, underrepresented structures. AlphaFold 3 represents a powerful step toward in silico aptamer design and offers a promising glimpse into a future where artificial intelligence accelerates discoveries and advancements in aptamers as effective affinity reagents.
