Progress in Structure Prediction and Mutation Visualization

Tom Goddard
June 18, 2026

• OpenFold 3, Boltz 3, AlphaFold 3 structure prediction.
• Combining deep mutation scan data with evolutionary sequence alignments and SNP data.
• Potential collaborations
• Miscellaneous

Structure prediction projects

These projects are exploring potential new capabilities in ChimeraX structure prediction.

Pooled AlphaFold3 to find protein-protein interactions Setup AlphaFold3 pooled predictions for finding protein-protein interactions with postdoc Horia Todor. Horia is in Carol Gross's lab, UCSF Dept Cell & Tissue Biology, research on virology & microbial pathogenesis. Run single predictions with 10 - 20 proteins from a microbial genome to find which interact. Horia runs batches of all genes in small microbial pathogen genomes. Paper Predicting the protein interaction landscape of a free-living bacterium with pooled-AlphaFold3 Setup on cloud GPU service Runpod.io to use Nvidia B200 (180 GB, $5.89/hour) and B300 (288 GB, $7.39/hour).

Prediction with 18 proteins. PAE plot at right shows 3 potential interactions (off-diagonal blocks).

Openfold 3 improvements by OpenFold team OpenFold 3 github has more developers, about 5 making commits now. Most commits are optimizing memory and speed, and simplifying installation. Not clear of private github still used, not seeing algorithm fixes, e.g. predictions without MSA are bad. Registered for Open Molecular Software Foundation workshop at UCSF Rutter Center on Wednesday June 24 which will have OpenFold demonstration.

Graph showing memory use optimizations of OpenFold. Can predict ~4800 residues with GPU with 80 GB.

Openfold predictions using ligands with explicit hydrogens "Fine-tuning" training of Openfold to use hydrogens on ligands Amit Elia from UCSF AI and Computational Drug Discovery (AICDD) will work this summer on this. We are discussing training using PDB ligands (chemical component dictionary). PDB validation reports check for idealized geometry that we may include in "loss function". If we train on ligands without protein, loss function needs to allow for rotatable bonds. Initial training studies using AICDD Nvidia RTX 6000 (96 GB) workstations. If successful try larger scale training on UCSF coreHPC cluster.

Cyclohexane predicted by OpenFold including hydrogens. Ring flattened. Colored by pLDDT. Cyclohexane from PubChem.

Computer hardware for running structure predictions

• Updated ChimeraX Boltz and OpenFold to newer Torch CUDA version 12.8 to work with Blackwell GPUs (e.g. Nvidia RTX 50x0 series) using Trooper.ai cloud GPUs.
• Fixed ChimeraX AlphaFold 2 predictions on Google Colab, broken by a tensorflow update on Colab.
• Helped Patrick Quinn (UCSF school of medicine) running AlphaFold 3 on Wynton.
• Klim Verba is arranging meeting with coreHPC admins to get students access for classes. Asked me to attend meeting to advise on coreHPC support for AlphaFold 3, Boltz, and OpenFold.
• Nvidia DGX Spark AI computer $5000 with 128 GB unified memory, GB10 GPU, for machine learning uses that need lots of GPU memory. Relatively slow with speed similar to Nvidia RTX 3090.
• Boltz developer Gabriele Corso asked if we could add ChimeraX interface to boltz.bio cloud prediction service. Cost per prediction $0.02 - $0.20 for 256 - 2048 residues.

DGX Spark has MacMini size

DGX Spark hardware

Combining diverse protein mutation data

Visualize protein variants (amino acid mutations) of human mu opioid receptor from multiple experimental data sources:

• Evolutionary sequence alignments across many species.
• Single nucleotide polymorphisms (SNPs) from about 800,000 human genome sequences.
• Deep mutational scan assays for activity with different opioid drugs.

Evolutionary sequence alignments combined with mutant activity scores

• Sequence search UniRef100 (475 million protein sequences) for similarity to human mu opioid receptor using mmseqs2 finds 400,000 sequences (E-value cutoff 1e-3).
• ChimeraX Profile Grid displays frequency of each amino acid at each residue position.
• Color profile grid for mutations that cause loss of function (red) or gain of function (green) in deep mutational scan with DAMGO ligand (analog of human endogenous opioid met-enkephalin).

How are loss of function mutations rescued?

• Willow is interested in mutations of residue K100.
• Evolutionary sequences show mutation K100H in 6% of sequences.
• What co-occuring mutations are present in K100H sequences?
• Click K100H in ChimeraX profile grid (orange diamond in middle) and color other mutations by change in frequency between K100H sequences versus all sequences. Orange 2x increase, White no change, Blue decrease.
• Color mu opioid receptor structure PDB 8ef5 residues with largest prevalence change for K100H red.

Clinical variants found in human mu opioid receptor

• UniProt aggregates genomic mutations from databases dbSNP, ClinVar, gnomAD, TCGA, BioMuta.
• UniProt Variant Viewer web page shows variants and possible disease associations.
• Human mu opioid receptor UniProt Variant Viewer

Visualizing clinical mutations with activity measurements

• Fetch Uniprot variants in ChimeraX: "open P35372 from uniprot_variants".
• Open deep mutational scan activity measurements from Willow: "open composite_dms_scores_R.csv".
• Create ChimeraX heatmap of clinical variants colored by experimental activity with DAMGO ligand from Willow's deep mutational scan. Red is loss of activity, blue gain of activity, white no change in activity.
• Willow believes the loss-of-function variants are not fatal because other mu opioid receptor allele is not mutated and one copy is sufficient.

Next steps in mutation visualization

• Allow fetch from Multiplexed Assays of Variant Effect database (MaveDB). I made ChimeraX read MaveDB .csv files.
• Look at the few hundred sequences in UniRef100 differing from human mu opioid receptor by 1 - 10 mutations. I've written some initial phylogenetic tree code, most close sequences from primates. Combine with activity scores to find compensatory mutations (not additive mutation effects).
• Tony Capra interested in combining deep sequence alignments, human SNPs, and DMS assays to understand evolution and human genetic diversity.
• Willow has new cystic fibrosis protein deep mutational scan data to help understand drug efficacy.
• Willow has olfactory receptor multi-variant (up to 10 variants) DMS data that will require new visualization methods to understand mutation evolution landscape.
• Bob Stroud interested deep sequence alignments and deep mutational scan for understanding GPCRs.

Other Potential Collaborations

• Joe DeRisi and graduate student Owen Moosman showed me antibody epitope mapping of picornaviruses possibly associated with type 1 diabetes, with web browser and ChimeraX visualization.
• Dyche Mullins and postdoc Kyle DeMarr want to use Chimerax microscopy visualization for studying butterfly wing pigments. Will meet in July to discuss details.

ProNA3D user ChimeraX interfaces.

Miscellaneous

• Postponed submitting R50 grant proposal until Dec 4, 2026. Topics and collaborators, Wah (cryoET), Willow (mutations), Aashish (mutations), David Fay (structure prediction).
• ChimeraX 1.12 released June 11, 2026. Tristan is working on ISOLDE release this week or next.
• Approved ProNA3D Toolshed bundle for RNA contact analysis combining cryoEM, AlphaFold3, and 2D graph visualization methods. BioRxiv preprint ProNA3D: Distance-Based Analysis of Nucleic Acid-Containing Interfaces