We’ve been testing the newly released ChatGPT5 in Rush as an experiment. Here’s a comparison of GPT5 and Grok’s response to the query ‘run a virtual screen for binders to BACE1’.

In summary:

• GPT5 is MUCH more verbose, and includes a LOT more technical detail.
• Some of this is useful - for example, it did a better job of conveying relevant information from our docs about how Rush’s computational tools work behind the scenes.
• Some of this is not useful - primarily the fact that it hallucinated a lot of untrue things about our platform and the available run settings.

We’ll be experimenting with different prompts and settings for GPT5, and seeing if it’s better for complex reasoning - but in the meantime we’ll be sticking to our regular models in Rush.

Grok Response

I've drafted a virtual screening run for binders to BACE1 (Beta-secretase 1), a critical target for Alzheimer's disease therapeutics due to its role in amyloid precursor protein cleavage. The run has been set up with Bayesian optimization to enhance screening efficiency, as per our default approach on the Rush platform. Below, I provide detailed information on the structural features of BACE1, the rationale for PDB selection, binding site residues, chain selection, and run configuration. You can review and finalize the configuration in the panel below.

Structural Features of BACE1 for Inhibitor Design

1. Catalytic Dyad: BACE1, an aspartyl protease, features a catalytic dyad of Asp32 and Asp228, essential for hydrolyzing amyloid precursor protein (APP) into β-amyloid peptides. Inhibitors typically target these residues via hydrogen bonds or electrostatic interactions, often using chemical groups like -NH2 for enhanced potency and specificity.References: DOI: 10.1186/1472-6807-12-21; [DOI: 10.1016/j.ejmech.2016.08.028.](https://doi.org/10.1016/j.ejmech.2016.08.028.)
2. Binding Subsites: The active site cleft includes subsites S1 to S4 and S1' to S4', with specific residues enabling diverse interactions:
   • S1: Leu30, Phe108, Ile118, Trp115 (hydrophobic).
   • S1': Thr72, Thr231, Thr232, Ile226, Val332, Asp228 (hydrophilic/hydrophobic).
   • S2': Tyr71, Pro70, Val69, Ile126, Tyr198 (amphipathic).
   • S2 and S4: Lys9, Thr72, Gln73, Arg235, Arg307 (hydrophilic).
   • S3' and S4': Pro70, Glu125, Arg128, Trp197 (mixed). Inhibitors spanning multiple subsites often achieve higher binding affinity.References: [DOI: 10.1016/j.jmgm.2017.06.001.](https://doi.org/10.1016/j.jmgm.2017.06.001.)
3. Flap Dynamics: A β-hairpin loop (residues 67-75) regulates active site access, with Tyr71-Trp76 hydrogen bonding stabilizing the closed conformation for enhanced inhibitor binding.References: [DOI: 10.3390/molecules27144372.](https://doi.org/10.3390/molecules27144372.)
4. 10s Loop: Near the S3 pocket, the 10s loop (including Ser10) adopts open conformations for access, with ligand binding inducing closure for stabilization.References: [DOI: 10.21203/rs.3.rs-3341477/v1.](https://doi.org/10.21203/rs.3.rs-3341477/v1.)
5. Selectivity: Structural differences, such as Tyr119 and the Pro118-Gln121 loop in BACE1 (versus Tyr87 in BACE2 and Tyr78 in Cathepsin D), are critical for designing selective inhibitors to avoid off-target effects.References: [DOI: 10.1016/j.molstruc.2020.127893.](https://doi.org/10.1016/j.molstruc.2020.127893.)

PDB Selection and Validation for Virtual Screening

Since no specific PDB ID was provided, I conducted a search for human BACE1 structures with resolutions better than 2.5 Å, bound to inhibitors, and covering the catalytic domain. Five structures were evaluated:

• 2G94 (1.86 Å, selective inhibitor ZPQ).