Synthetic sequence alignments as programmable probes of learned conformational landscapes in deep learning protein structure predictors

Synthetic sequence alignments as programmable probes of learned conformational landscapes in deep learning protein structure predictors Abstract What deep learning protein structure predictors learn about conformational landscapes remains largely opaque. Here we introduce synthetic multiple sequence alignments (MSAs), designed by inverse folding to encode predefined structural constraints, as a programmable intervention for interrogating the internal logic of structure prediction systems. Synthetic MSAs systematically bias AlphaFold2, AlphaFold3, and RoseTTAFold2 toward distinct conformational states of fold-switching proteins, including alternative conformations inaccessible through natural sequence information alone. Adversarial experiments pairing query sequences with MSAs encoding competing folds reveal sequence-dependent responses, exposing how alignment-derived and sequence-derived signals are weighted within each system. Probing predictions initialized from molecular dynamics trajectories reveals a systematic bias toward compact, training-distribution-favored conformations. Hybrid alignments combining synthetic and natural MSA segments enable targeted steering toward specific conformational states. These results suggest synthetic MSAs as a generalizable framework for dissecting the conformational landscapes encoded by deep learning structure predictors, with direct implications for understanding model behavior and accessing biologically relevant hidden states. Content adversarial --> structures of adversarial analysis Fast_folding are from fast-folding proteins Negative_examples and New_examples are both "diverse proteins" ablation tests previously working MSAs with previously stubborn sequences fast_folding_analysis --> use our pipeline on sampled frames from molecular dynamics simulations _unfolding --> high-temperature unfolding simulations saved in another folder here no suffix --> the long simulations can be downloaded from this paper FrankenMSA --> MSAs and predictions of the bombyx mori and RSV trimer protein porter_dataset --> protein structure predictions with MMseqs2 and synthetic MSAs AF-cluster predictions were omitted as they were too many and did not fit in the data limit porter_simulations --> MD simulations of unrecovered proteins from Extant fold-switching proteins are widespread test_set_structures --> AlphaFold2 testing of synthetic MSAs compared to MMseqs2 MSAs of four proteins that were released after the training set has been collected. SA1 from AlphaFold predictions of fold-switched conformations are driven by structure memorization and ASCT2, STP10, ZNT8 from Sampling alternative conformational states of transporters and receptors with AlphaFold2 (MCT1 was discarded because the TM-score of the two modeled regions was too minimal between the two conformations) Contact If you have questions, please check our GitHub or otherwise contact [e-mail] or [e-mail].