Conditioning

How runtime inputs such as MSAs, homolog prompts, templates, ligands, and initial structures affect predictions. Active reward-driven changes to conditioning also belong under inference/guidance.

Quartz 4

Explorer

Conditioning

Surface-complementary seeds expand the structural diversity of diffusion-designed protein binders

Coevolutionary patterns in multiple sequence alignments do not contribute to protein-protein complex prediction

AlphaFold3 inference benefits more from phylogenetically diverse MSAs than deep redundant MSAs

MSAs make AlphaFold3 predictions robust to extensive mutation and deletion

Antigen context improves CDRH3 structure prediction

Alternate conformations can be sampled with MSA-based structure prediction methods using custom PDB databases and subsampled MSAs

Inclusion of adjacent loops can improve CDRH3 structural modeling

Multi-sequence protein language models

Number of sequences with 90% homology is a greater predictor of viral variant effect prediction than alignment depth

Scrambled MSAs outperform single-sequence-prediction with AlphaFold

Structure prediction with AlphaFold improved when starting from an initial guess, rather than default initialization

Variant effect prediction with homology-aware PLMs improves with ensembling of multiple prompts

pLDDT correlates with number of homologous sequences provided during runtime

Conformational sampling by AlphaFold3-generation methods can be achieved by scaling the pair representation

De novo structure prediction with ligands outperforms apo de novo structure prediction alone