User:Eric Martz/AlphaFold3 case studies

From Proteopedia

Current revision

Summary:
The purpose of these case studies was for me (User:Eric Martz) to gain experience using the AlphaFold3 Server. See resulting guidance for using the AlphaFold3 server.

Comparisons were made between 3 empirical models and predictions by the AlphaFold3 Server and the AlphaFold2 Database. Predictions by AlphaFold3 had high confidence (except for disordered regions), and high accuracy for naturally occurring sequences, including a 3-chain assembly with complex structural inter-dependencies (1a0r). In the two cases checked, the predictions of AlphaFold2 were slightly less accurate than those of AlphaFold3.

In the case of chain A of 8jrp, loop 397-423 is predicted to be disordered, but nevertheless has coordinates substantiated by the cryo-EM density map. AlphaFold3 had no confidence in its prediction of this segment, and placed it in a different position than in the empirical model.

One synthetic construct was analyzed (8k8g), for which no similar sequences or structures are in the PDB prior to 2024^[1]. AlphaFold3's stated training set excludes PDB entries after September, 2021. AlphaFold3 had no confidence in its prediction for the fold of this chain, which was very different from the empirical structure. The empirical model is a homodimer, but none of the 5 homodimers predicted by AlphaFold3 (with no confidence, ipTM 0.10-0.37) had inter-chain contacts similar to the that in the empirical structure.

AlphaFold3 Example 8AW3

Summary: This is one of three examples featured in the AlphaFold3 Server. AlphaFold3 predicted the fold of the mid-341 residues of tRNA A34 deaminase (UniProt Q381Q7) accurately, when compared to the empirical model in 8aw3. It made no confident predictions for what appear to be 90 residues in two disordered loops, and a 25-residue disordered C terminus. In contrast, the prediction from the AlphaFold2 Database was bent relative to the empirical model, but otherwise largely accurate.

UniProt Q381Q7 (tRNA A34 deaminase) is the longest chain in the example 8aw3 provided by the AlphaFold3 Server. In 8aw3, its chain name is "3". 8aw3 is a cryo-EM structure with resolution 3.6 Å. 8aw3 indicates that the experimental material included full-length chain 3 of 369 residues, but 28 end-residues lack coordinates, so the coordinates run from UniProt 4-344, length 341. Within this range, there are two loops lacking coordinates of lengths 49 and 13, leaving 279 amino acids with coordinates. Neither missing loop is predicted to be disordered at UniProt, although the estimated disorder propensity is high for both.

Q381Q7: AlphaFold3 vs Empirical Structure

The overall 4-chain prediction (for the complex in 8aw3) has pTM of 0.75, at the low end of the range 0.7-0.9 deemed "confident" by the server. Both of the missing loops have "no-confidence" (red) coordinates. The longer one is modeled as a large circle protruding from the surface of the protein, typical for loops with intrinsic disorder. The average pLDDT for chain 3 is 68 (weak confidence), but with the large disordered loop removed, this increases to 74^[2].

The AlphaFold3 prediction superposes (FATCAT rigid) with all 279 alpha carbons of chain 3 of 8aw3 with RMSD 2.2 Å. The morph shows the largest discrepancy in the red (no confidence) loop at center bottom, which has sequence SNSGCRKSNR (238-247 in the numbering of both UniProt and 8aw3). These residues have coordinates in 8aw3, but 3 residues, CRK, have incomplete sidechains.

Q381Q7: AlphaFold3 vs AlphaFold2 Database

8JRP Chain A: Partially Untemplated

Missing/disordered

N-terminal residues 1-119, missing in 8jrp, are predicted to be partially disordered. Loop 170-230, missing from 8jrp, is predicted to be disordered at RCSB. Sequence range 388-424 is predicted to be disordered, but has coordinates in 8jrp.

AlphaFold 3 prediction

Chain A in 8jrp is human ubiquitin-protein ligase E3A. The full-length sequence, Q05086, has 875 amino acids. The 3.6 Å resolution cryo-EM structure indicates that the full-length protein was imaged, but the N-terminal end of the model lacks coordinates for 119 residues. The ends of the model are sequence numbers 120 and 869 (length 750), but one loop of length 61 (170-230) is missing coordinates, leaving 689 amino acids with coordinates. Conveniently, the sequence numbers in UniProt and 8jrp are the same.

AlphaFold3 was unable to predict with confidence the loop 170-230 (length 61) missing in 8jrp, and predicted to be disordered. The average pLDDT for the predicted loop was 40.2 (no confidence). Similarly, the N-terminal residues 1-119 missing from 8jrp were predicted with average pLDDT of 55.5 (no confidence). The 27-residue loop 397-423, predicted to be disordered but with coordinates in 8jrp, appears red in the predicted model with average pLDDT 47.1 (no confidence). In contrast, the average pLDDT for the residues having coordinates in 8jrp (120-169, 231-869) is 86.1 (high confidence).

The AlphaFold3 prediction superposed 688/689 alpha carbons onto chain A of 8jrp with RMSD 2.1 Å. Thus, the prediction is very close to the empirical model. The above morph shows that the largest discrepancy between the prediction and the empirical model occurred in the 27-residue loop 397-423, predicted to be disordered. These residues have coordinates in 8jrp, and the EM density map shows strong evidence for their positioning in the empirical model.

Templated vs. Untemplated

The apparently untemplated coordinates in chain A of 8jrp are 120-519. This range is missing coordinates for the 61-residue disordered loop. After removing coordinates for the 37-residue loop that is predicted to be disordered, but which has coordinates, there are 302 residues remaining. Superimposing these 302 alpha carbons with the AlphaFold3 prediction gave RMSD 0.9 Å. Thus, the prediction for the putatively untemplated half was excellent.

The templated coordinates are 520-869, length 350. Superposition gave RMSD 1.3 Å. In conclusion, the putatively untemplated half was predicted as well as the templated half.

Five AlphaFold3 models

The AlphaFold3 server delivers five models. In this instance, four all have pTM 0.77, and one (model_3.cif) had pTM 0.76. Superposing the latter with the empirical model gave RMSD 2.1 Å for all 689 alpha carbons in the empirical model. This is the same RMSD as obtained with the "top" model (model_0.cif). In conclusion, there was no significant difference between the abilities of these two of the five models to superpose with the empirical model.

AlphaFold2 Database prediction

A rigid superposition of the AlphaFold2 Database prediction for Q05086 with chain A of 8jrp gave RMSD 3.0 Å. The main discrepancy, aside from that expected for the loop 397-423 (red, no confidence), was rotation of the C-terminal domain. Yet when submitted allowing twists, no improvement occurred. In conclusion, the AlphaFold3 prediction was better than the AlphaFold2 prediction.

1a0r: Three Sequence-Distinct Chains

Empirical Model 1a0r

AlphaFold3 Prediction for 1a0r

Untemplated Synthetic Construct 8K8G

Superposition RMSD Summary

Each superposition in this table is discussed above.

AF3 = AlphaFold3 Server. Model 0 (best of five, 0-4) was used unless otherwise noted.
AF2 DB = AlphaFold2 DataBase
Note a: Disordered loop 57-108 deleted from both models.
Note b: PyMOL super command, limited to alpha carbons, e.g. super empirical////ca, af3////ca.
Note c: ChimeraX Tools, Structure Analysis, MatchMaker. Default "pruning" eliminated pairs farther than 2.0 Å apart. Also did a cutoff of 1.0 Å.

Methods

These case studies of predictions by the AlphaFold3 Server were done in November 2024. At that time, the AlphaFold Database had been generated with AlphaFold2. Superpositions and RMSD values for pairs of chains were obtained from FATCAT. Superpositions and RMSD values for multiple-chain assemblies were obtained from PyMOL with its super command using default settings, and/or with ChimeraX using its Tools, Structure Analysis, MatchMaker, since FATCAT and the RCSB Superposition utility are limited to pairs of chains. Rocking animations were generated by FirstGlance in Jmol, and morphs were captured from FATCAT. pLDDT averages for specified sequence ranges were obtained using FirstGlance in Jmol, after converting predicted .cif files to .pdb format.

See Also

• AlphaFold3 Server alphafoldserver.com
• How to predict structures with AlphaFold
• AlphaFold/Index lists pages in Proteopedia that relate to AlphaFold.

Notes

1. ↑ For the sequence in 8k8g, the top BLAST hit in UniProt is 37% identical for 79% of length.
2. ↑ Average pLDDT values were obtained with FirstGlance in Jmol, which reports the minimum, average, and maximum in the upper right as "Reliability".
3. ↑ 8jrp was chosen despite its poor resolution of 3.6 Å because the better-resolution 8jrn, 2.6 Å, is not available in PDB format. A rigid superposition by FATCAT matched all 689 residues with coordinates, giving RMSD 0.7 Å, a near perfect match.
4. ↑ A FATCAT rigid superposition of 8jrp chain A with 1cz4 chain C shows a twist between the two domains. A FATCAT flexible superposition with one twist gives RMSD 1.3 Å for all 350 alpha carbons. Thus, 1c4z would be an excellent template for predicting the C-terminal half of 8jrp chain A.
5. ↑ Predictions were done for the full length Q05086 sequence (shown above), and with 1-119 removed. Rigid FATCAT superposition between the two predictions (only the 689 alpha carbons present in 8jrp) gave RMSD 1.4 Å. Thus, inclusion of 1-119 had little if any effect on the prediction of the remainder of the structure.
6. ↑ 8kck: Rfree worse than average. PDB-ReDo increases both R and Rfree! 8k8g: PDB-ReDo reduces R from 0.20 to 0.17, and Rfree from 0.222 ("unreliable") to 0.207 ("worse than average"). Characterizations "worse than average" and "unreliable" are from FirstGlance, which uses statistics from the PDB.