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-toend 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.

Systematic identification of conditionally folded intrinsically disordered regions by AlphaFold2

“…We hypothesized that the predicted IDRs with high pLDDT scores might manifest for one or a combination of reasons: (1) global amino-acid sequence differences in comparison to the predicted IDRs with low pLDDT scores, (2) relatively high positional sequence conservation (i.e., “high quality” multiple sequence alignments (MSA)), and (3) the enrichment of high-pLDDT IDR sequences in the PDB. The first possibility would reflect a differential “folding propensity” that is inherently encoded in the amino-acid sequences of high vs. low pLDDT-scoring IDRs, whereas the latter two possibilities would influence the AlphaFold2 prediction confidence due to the depth of the MSAs (2) or sequence similarity to the structures from the PDB used in training (3) (Jumper et al 2021a,b). Given the relatively poor coverage of IDRs in the PDB (Quaglia et al 2021) and the poor positional alignability for most IDRs (Colak et al 2013; Nguyen Ba et al 2012; Zarin et al 2019, 2021), it is plausible that some combination of all three of the aforementioned possibilities could contribute to high pLDDT scoring IDRs.…”

“…The biennial Critical Assessment of Structure Prediction (CASP) competition (Moult et al 1995) has stimulated many developments in the field of protein structure prediction, including the successful implementation of co-evolutionary restraints derived from multiple sequence alignments (MSAs) and machine learning protocols in CASP12 (Moult et al 2018; Schaarschmidt et al 2018). CASP14 brought a revolutionary advancement: the AlphaFold2 team at DeepMind produced more models with atomic-level accuracy than ever before in the history of CASP (AlQuraishi 2021; Jumper et al 2021a,b). The second-best scoring prediction software in CASP14 led to the RoseTTAFold structure prediction platform, which was released in open-source format and contained a webserver for ease of access (Baek et al 2021).…”

“…Here, we show that thousands of IDRs are predicted by AlphaFold2 to be folded with high (70 ≤ x < 90) or very high (≥ 90) predicted local difference distance test (pLDDT) scores (Mariani et al 2013), which measure the confidence in the predicted structures (Jumper et al 2021b). We find that, compared to IDRs with low pLDDT scores, the amino-acid sequences of IDRs with high pLDDT scores are enriched in charged and hydrophobic residues, show more positional conservation, and have more alignment matches to sequences in the PDB.…”

Predicting direct physical interactions in multimeric proteins with deep learning

“…Moreover, only one of the five multimer DL models generated such a model by using unpaired MSAs. It appears that using unpaired MSAs may be the key for generating this good model, because runs with paired MSAs by following the AF-Multimer workflow return unphysical models with severe clashes, a phenomenon akin to “chain collapse” observed previously[3]. Paired MSAs may have contributed to the issue.…”

AlphaFold Models of Small Proteins Rival the Accuracy of Solution NMR Structures

“…In the 2020 Critical Assessment of Protein Structure Prediction (CASP14), the DeepMind AlphaFold2 (AF2) deep learning method (Jumper et al, 2021a; Jumper et al, 2021b) demonstrated outstanding performance in blind predictions of protein structure, delivering excellent structural matches to experimental models derived from X-ray crystallography, NMR and cryoEM data, over a wide range of target difficulty (Kryshtafovych et al, 2021). These AlphaFold2 model predictions had an unprecedented high accuracy, assessed by backbone atomic coordinate global distance test (GDT_TS) scores.…”