Protein kinases are critical drug targets, requiring therapeutics that can modulate their active and inactive conformational states. While cofolding models can generate global folds directly from kinase sequences and ligand SMILES strings, these models have not yet been tested on their ability to recover ligand-induced-fit conformational states of the kinase proteins. Here, we introduce KinConfBench, a curated benchmark of 2225 high-quality human kinase chains to evaluate the ability of four state-of-the-art cofolding models—Boltz-2, Chai-1, Protenix, and RoseTTAFold-All-Atom—to recover both canonical and rare conformational states. We show that geometric success metrics of a ligand pose in the active site do not correlate strongly with the correct kinase conformational state, motivating a new set of dynamical benchmarks for assessing cofolding models. While all four cofolding models achieve ~60–80% prediction accuracy for kinase conformational classification, they exhibit severe mode collapse when performing multiple inferences, show negligible structural diversity in sampling induced-fit motions, and display a prevalent “apo-drift” in which most cofolding models predominantly predict the kinase to be in its ligand-free state. Our results highlight that capturing ligand-induced protein conformational diversity, not just geometric fit, is critical for next-generation structure-based drug discovery.
Kunyang Sun, T. Head-Gordon· npj Drug Discovery· 0 citations
Co-folding models hold immense potential for allosteric drug discovery, but have been severely hampered by their systematic bias toward orthosteric ligand binding. While fragment screening has been proposed for allosteric binding site discovery, we show that co-folding models still suffer from memorization in which chemically simpler fragments also default to canonical orthosteric binding sites. To overcome these limitations, we introduce CAFE (Co-folding Approach for Fragment Exploration), a co-folding protocol that uses competitive orthosteric blockers to divert fragments into non-canonical sites as illustrated here with the Boltz-2 co-folding model. Using ADP as an orthosteric blocker for the kinase family, we find CAFE substantially increases the allosteric binding site exploration for fragments, with notably strong absolute binding free energies that match or exceed those of known crystallographic poses, without post-hoc refinement of the Boltz-2 prediction. We also show that CAFE identifies cryptic binding pockets undetected by conventional pocket prediction tools, some of which are more thermodynamically favorable than the allosteric or orthosteric pockets. To demonstrate generality, we apply CAFE using Type I orthosteric blockers for kinase proteins, known orthosteric ligands as blockers for non-kinase proteins in the RAS-MAPK signaling pathway, and for virtual screening campaigns using fragment libraries for new fragments that selectively engage allosteric and cryptic binding sites. CAFE establishes orthosteric blocking and fragment screening as a training-free, inference-time protocol that helps overcome some of the limitations of current co-folding models while elevating their great promise for allosteric and cryptic binding drug discovery.
Justin Purnomo, Kunyang Sun, T. Head-Gordon· bioRxiv· 0 citations
It is shown that enhancing an LLMs capabilities for robust prediction of small molecule geometries still retains nearly all of its pre-trained language abilities by randomly mixing in small quantities of natural language prompt-response pairs into the fine-tuning.
Joe Cavanagh, Jonathan Arnold, G. Alteri et al.· 0 citations