Work in progress September 20 - November 20, 2024

Tom Goddard
November 21, 2024
Group meeting

Imatinib boound to ABCG2 PDB 6vxh

DOX in DNA PDB 1d12. MTX in DNA PDB 4g0v

Topics

• Demonstration of deep mutational scan visualization.
• Initial experience running AlphaFold 3 locally, ligand binding predictions

Deep mutational scan of human transporter ABCG2

• High throughput method to test every possible single amino acid mutation in a protein.
• Example human ABCG2 transporter protein pumps drugs out of cells.
• 655 residues * 19 mutations/residue = 12445 mutations.
• Multiple assays score how cells function with each mutation.
• Test efflux of chemotherapy drugs, topoisomerase inhibitors that kill cancer cells by blocking unwinding of DNA, DOX (doxorubicin), MTX (mitoxantrone), SN38 (irinotecan).
• Also test how much ABCG2 makes it into membranes for all mutants.

Demonstration of deep mutational scan visualization

DOX vs MTX scatter plot. Synonymous blue, Mutations near R482 green.

R482 mutations have high DOX efflux.

Mutations of R482, S440, S443 have high DOX efflux.

• Open mutations file abcg2.csv.
• Show dox and mtx on x and y axes of scatter plot.
• Open ABCG2 structure 6vxf, color by chain.
• Click right-most (highest dox efflux) mutation and choose "Color mutations near residue R482".
• Click R482 and nearby yellow and orange mutations and "Show side chain" and "Color to match plot".

Activating mutations near drug ligands

ABCG2 with imatinib bound, PDB 6xvh.

ABCG2 with imatinib, hbonds broken.

ABCG2 with SN38 bound, hbonds rearranged.

• In ABCG2 bound to imatinib hydrogen bonds are broken.
• Serines 440 and 443 are close to the imatinib ligand.
• May explain why R482 specifically increases DOX efflux causing chemotherapy drug resistance.

AlphaFold 3 code release

• Google released AlphaFold 3 code last Monday, November 11, almost exactly 6 months after they said they would.
• Code is only for non-commercial use.
• Must fill out a form to get the trained neural net parameters from Google, takes 2-3 days to get "approved".
• Released code can predict arbitrary ligands with proteins and nucleic acids and modfied residues including glycosylations, and also ions.
• Officially Google only support Nvidia H100 80 GB and A100 80 GB GPUs ($20000 - $30000 each), but tests show it runs on Nvidia RTX 30x0 and 40x0 (24 GB) but not on RTX 20x0 or older because it requires Nvidia compute capability 8.0.

Initial AlphaFold 3 test runs on Wynton and Quillian

• I've made about 8 predictions with proteins (lengths 100-600 amino acids) and ligands.
• A few predictions failed on Quillian during sequence alignment calculation apparently took more than 64 Gbytes.
• Predicted influenza neuraminidase with two zanamivir antiviral drug molecules (PDB 2cml). Got protein right and one of 2 zanamivirs right. 389 residues, 23 atoms per zanamivir, took 30 minutes on Wynton (A40 48GB GPU) almost all sequence alignment calculation, with 76 seconds for GPU structure prediction.

Influenza neuraminadase (tan) and Zanamivir (orange) PDB 2cml, AlphaFold 3 (blue) with Zanamivir (pink).

One Zanamivir placed correctly and one incorrectly.

Correct Zanamivir and all side chains match.

Predicted aligned error, correct and incorrect zanamivirs in lower right corner.

Nipah virus plus Zanamivir antiviral drug AlphaFold 3 prediction

• Predicted Nipah virus (PDB 8xps) with two zanamivir antiviral drug molecules (no experimental structures with this antiviral). Got protein right and one of 2 zanamivirs in approximately the homologous position from influenza neuraminadase.
• Nipah G-protein and influenza neuraminadase have remote similarity fund by Foldseek.
• 602 residues, 23 atoms per zanamivir, took 26 minutes on Wynton (A40 48GB GPU), just 125 seconds for structure inference.

Nipah G-protein AlphaFold 3 (blue) aligns to PDB 8xps (pink).

AlphaFold 3 Nipah G-protein (blue) with zanamivir (pink) aligned to Influenza neuraminadase (tan) and zanamivir (orange) PDB 2cml

Two zanamivirs in AlphaFold (light and dark pink) are stacked.

Predicted aligned error, nonmatching and matching zanamivirs in lower right corner.

AlphaFold 3 next steps

• Figure out how to do compute fast sequence alignments (MSA) so prediction can run 10x faster.
• Compare sequence databases in RAM-disk, or NVMe drive, or SSD drive, or network file system.
• Do MSA calculation in CPU job to avoid wasting GPU resources.
• Figure out how to do MSA calculation with 64 GB of memory.
• Write web page to show UCSF users how to run AlphaFold 3 jobs. Use their own model parameters obtained from Google.