The transformative power of transformers in protein structure prediction

Abstract

Transformer neural networks have revolutionized structural biology with the ability to predict protein structures at unprecedented high accuracy. Here, we report the predictive modeling performance of the state-of-the-art protein structure prediction methods built on transformers for 69 protein targets from the recently concluded 15th Critical Assessment of Structure Prediction (CASP15) challenge. Our study shows the power of transformers in protein structure modeling and highlights future areas of improvement.

Keywords: deep learning; neural networks; protein structure prediction; transformers.

Fig. 1.

Performance benchmarking of the state-of-the-art protein structure prediction methods on the CASP15 dataset. (A) GDT-TS score comparisons. The dashed lines represent the mean performance with the percentages reported in the Top Left indicating that the method in the y axis outperforms the method in the x axis, and vice versa for the Bottom Right. (B) Grishin plot analysis. (C) Domain-level GDT-TS scores against the length of the domains with the inset showing the domains having lengths less than 750 residues. The lines represent linear fit to the data. (D) Three representative CASP15 targets with the predictions (in rainbow) superimposed on the experimental structures (in gray). Bold numbers indicate the best performance. (E) Correct overall topology prediction performance in terms of %TM-score >0.5. (F) lDDT scores of MSA-based methods against the MSA depth measured by the logarithm of Neff. The lines represent linear fit to the data. (G) lDDT scores against the length of the protein targets with the inset showing the targets having lengths less than 750 residues. The lines represent linear fit to the data. (H) Ramachandran plot analysis. The color code ramps from blue to red for low to high density. (I) MolProbity score distributions. (J) GDC-SC score comparisons similar to A.