Applying and improving AlphaFold at CASP14

Abstract

We describe the operation and improvement of AlphaFold, the system that was entered by the team AlphaFold2 to the "human" category in the 14th Critical Assessment of Protein Structure Prediction (CASP14). The AlphaFold system entered in CASP14 is entirely different to the one entered in CASP13. It used a novel end-to-end deep neural network trained to produce protein structures from amino acid sequence, multiple sequence alignments, and homologous proteins. In the assessors' ranking by summed z scores (>2.0), AlphaFold scored 244.0 compared to 90.8 by the next best group. The predictions made by AlphaFold had a median domain GDT_TS of 92.4; this is the first time that this level of average accuracy has been achieved during CASP, especially on the more difficult Free Modeling targets, and represents a significant improvement in the state of the art in protein structure prediction. We reported how AlphaFold was run as a human team during CASP14 and improved such that it now achieves an equivalent level of performance without intervention, opening the door to highly accurate large-scale structure prediction.

Keywords: AlphaFold; CASP; deep learning; machine learning; protein structure prediction.

FIGURE 1

Examples of visualizations used in the prediction checking Colaboratory notebook, shown with CASP target T1101

FIGURE 2

T1024: (A) Per‐residue lDDT‐Cα and pLDDT of T1024. Vertical gray shading indicates residues missing in the experimental structure, and colored shading indicates minimum and maximum values over five predictions. The pLDDT shows low confidence in the linker region indicating possible flexibility and qualitatively agrees with the true per‐residue lDDT‐Cα. (B) Unrealized distances in the expected distances of T1024 indicating possible alternate relative conformations of the two domains

FIGURE 3

T1044: Comparison of the number of effective alignments (Neff) per residue for each MSA, derived both from domain sequences and from cropping the full sequence MSA. Four domains (T1033, T1039, T1040, and T1043) substantially benefit from using the full sequence MSA. The dashed green line shows the approximate 30 alignment threshold considered sufficient for a good prediction with AlphaFold

FIGURE 4

Comparison of three different prediction methods for the targets with significant interventions: “Original system” is the automated prediction system as it existed at target release. “Submitted prediction” is the submitted structure prediction. “Final system” is the automated system as it existed at the end of the CASP14 assessment, improved by experience

FIGURE 5

All models produced for T1064–pLDDT versus final lDDT‐Cα. The strong correlation indicates that ranking many predictions by pLDDT was a successful strategy for this target