01丨Why Can High-Scoring Models Still Produce “Wrong Answers”?

Algorithmic Bias Amplification

The research team analyzed the widely adopted Crossdocked2020 dataset. Compared with experimentally resolved native crystal complexes, the Crossdocked dataset exhibits multiple prominent biases:

• The average molecule duplication count reaches 11.52, versus only 1.67 in crystal data;

• Artificial cross-docking pairing disrupts the native volume complementarity between pockets and ligands;

• Hydrogen-bond probability of ILE residues is overestimated by approximately 12%;

• Hydrogen bonds within the 2.75–3.5 Å range are substantially overpredicted;

• Key interactions such as halogen bonds are systematically underestimated.

Figure 1.Comparison between CrystalDataset and Crossdocked datasets in terms of molecular duplication, physicochemical properties, pocket matching, and interaction distribution. Figure sourced from the original paper.

Figure 2. Overview of the EIP-Diff framework, including global self-attention, interaction prompt embedding, prompt-guided cross-attention, and the diffusion generation process. Figure sourced from the original paper.

Cross-attention updates ligands only along the protein-to-ligand direction, enabling the generation process to explicitly respond to designated residues and interaction types.

• During training, interaction prompts act as explicit biological priors;

• During inference, interaction prompts serve as controllable instructions for molecular design.

03丨45,318 Crystal Complexes: Unleashing the Model’s Geometric Potential

Figure 3. Pharmacological property distribution, atomic composition distribution of different models, and generation success rates under varying numbers of interaction prompts. Figure sourced from the original paper.

04丨From Recapturing Binding Modes to Optimizing Novel Interactions

Aggregate metrics cannot fully determine whether a model truly understands a specific protein pocket. Accordingly, the team further selected KAT6A and YTHDC1 for structural validation and molecular optimization.

KAT6A：Recovering Native Binding Modes

In the KAT6A case (PDB: 6OIO), EIP-Diff achieves a 3D structural similarity score of 0.467, ranking first among six benchmarked models.

The generated molecules retain the acyl-sulfonylhydrazide core of the original ligand and form stable hydrogen bonds with key residues including Ser690, Arg655, Gly659 and Arg660. Hydrophobic substituents can effectively fill the pocket.

YTHDC1：Optimization While Preserving Critical Interactions

In the YTHDC1 case (PDB: 8K2E), the team adopted the key interactions of the reference ligand as prompts to generate and screen approximately 30,000 candidate molecules.

The top design maintains the critical orientation of the original carboxyl group and exhibits superior MM/PBSA binding free energy in molecular dynamics simulations:

•Designed molecule: −36.43 ± 1.41 kcal/mol；

•Reference molecule: −26.02 ± 1.08 kcal/mol。

The newly designed molecule additionally forms stable hydrogen bonds with SER362 and SER378.

Figure 4. Recapitulation of KAT6A binding modes and molecular optimization for YTHDC1, including structural similarity, binding conformations, binding free energies and interaction analysis of key residues. Figure sourced from the original paper.

05丨Toward Experimental Validation: IDO1 Candidate Molecule Achieves 0.31 nM

IDO1 is a vital target for tumor immunotherapy. Its inhibitor BMS-986205 undergoes a dynamic binding process from the surface-extended state via a bent transition state to a deep stable state, imposing stringent requirements on the 3D conformational capability of generative models.

For this complex pocket, the team anchored the chlorobenzene moiety at the pocket entrance and performed generative scaffold hopping on the key tail region, yielding 30,000 unique molecules.

Following multi-parameter screening including molecular docking, synthetic accessibility and QED, researchers finally selected two candidate compounds for experimental validation.

In the IFN-γ-stimulated HeLa cell assay, the IDO1 inhibitory activities of the two candidates are as follows:

•Candidate Compound 1: IC₅₀ = 0.75 nM;

•Candidate Compound 2: IC₅₀ = 0.31 nM；

•Positive control: IC₅₀ = 0.29 nM。

Figure 5. Scaffold hopping design for IDO1, candidate molecular structures, binding conformations, and cellular dose–response results. Figure sourced from the original paper.

06丨Where Do the Advantages of EIP-Diff Lie?

From Statistical Correlation to Biological Constraints

Instead of forcing the network to infer all binding rules independently, EIP-Diff introduces residue-level interactions as explicit conditions, aligning the generation logic of the model with real medicinal chemistry design principles.

Unlocking the Value of High-Quality Data

CrystalDataset provides not only an expanded set of samples, but also native pairings, experimentally determined conformations and authentic interaction annotations. Results reveal that high-fidelity data can only be translated into improved 3D geometric performance when supported by a suitable model architecture capable of utilizing such information.

Balancing Fidelity, Explorability and Controllability

Enabling Closed-Loop Experimental Validation

Conclusion

This work highlights not merely a 3D molecular generative model with enhanced benchmark performance, but a paradigm of co-design for architecture and datasets:

Leveraging experimental crystallographic data for reliable supervision, adopting SE(3)-equivariant networks to preserve 3D geometric consistency, and employing explicit interaction prompts to translate medicinal chemistry knowledge into actionable generative conditions.