DiffDecip

DiffDecip is a 3D target-aware diffusion model for lead optimization by scaffold decoration: given a protein pocket and a fixed molecular scaffold, it generates the R-group that completes the ligand. It was developed by groups at the School of Computer Science and Engineering and the School of Pharmaceutical Sciences at Sun Yat-sen University, with the West China School of Pharmacy at Sichuan University.

Its motivating observation concerns a gap in interaction-guided generative models. Diffusion models such as InterDiff, IPDiff, and IRDiff already steer generation toward molecules that form favorable non-covalent contacts inside a pocket, which raises predicted affinity. But contacts are not all equal: interactions with evolutionarily conserved residues are the ones that tend to underlie protein function and ligand bioactivity, and nothing in an affinity-driven objective distinguishes a conserved residue from an incidental one. DiffDecip adds that distinction explicitly, injecting per-residue conservation scores into the pocket representation and using a pretrained interaction network to shift generated atoms during both the forward diffusion and the reverse sampling process.

The model extends DiffDec, the same group's E(3)-equivariant conditional diffusion model for pocket-based scaffold decoration, and is evaluated directly against it. Among catalog entries it is closest in spirit to pocket-conditioned generators such as PocketXMol and Molexar, and stands in useful contrast to PharmaDiff, which conditions on a pharmacophore hypothesis and needs no protein structure at all.

#Key Features

• Conservation-aware pocket conditioning: Each pocket residue carries a conservation score in [0, 1] derived from a multiple sequence alignment, appended to the pocket descriptor so the denoiser can grow R-groups toward conserved regions.
• Interaction-prior guidance: A pretrained IPNet, built from SE(3)-equivariant networks and cross-attention and supervised on binding affinity, supplies interaction representations that guide atom coordinates rather than acting only as a post-hoc reranking score.
• Guidance in both diffusion and sampling: Conservation and interaction signals enter the forward noising process as fixed conditions and the reverse process as coordinate shifts, rather than being applied at sampling time alone.
• Scaffold-preserving generation: The scaffold and pocket are held fixed as context throughout, so the task is a genuine optimization of an existing lead rather than unconstrained de novo design.

#Technical Details

DiffDecip models the conditional distribution p(R | P, S) over R-group atom coordinates and types given pocket P and scaffold S, using a variance-preserving cosine noise schedule parameterized through the signal-to-noise ratio. Conservation scores are computed by retrieving each pocket's full sequence via its UniProt identifier and running HHblits against UniRef30_2023_02. Training data follow DiffDec's preparation of CrossDocked: from 22.5 million docked complexes, poses with RMSD below 1 Å and pocket sequence identity below 30% were retained, then the standard 100,000-pair training and 100-pair test split was sliced into scaffold and R-group by reaction-based rules, yielding 76,370 training tuples and 49 test tuples. Evaluation generates 100 molecules per test pocket. DiffDecip improves the mean QVina score to -8.19 kcal/mol from DiffDec's -8.11, raises the high-affinity rate — the share of cases beating the reference ligand — from 42.9% to 44.6%, and lifts uniqueness from 53.3% to 67.3%, at a small cost in validity (90.1% vs. 91.8%). Per target, it achieves the better binding affinity on 65.3% of the 49 pockets, forms more non-covalent interactions on 71.1%, and forms more interactions specifically with highly conserved residues (score above 0.4) on 84% — evidence that the conservation signal, not merely added interaction guidance, drives the gain.

#Applications

The setting is hit-to-lead and lead optimization, where a medicinal chemistry team has a validated scaffold in a known pocket and needs substituent proposals. Biasing those proposals toward conserved residues is chemically motivated beyond raw affinity: contacts at conserved positions are harder for a target to escape through mutation, which matters for antimicrobial, antiviral, and oncology programs where resistance is the failure mode. The conservation feature requires only a UniProt identifier and a standard MSA, so it can be computed for any target with a sequence.

#Impact

DiffDecip's contribution is the demonstration that evolutionary conservation is a usable conditioning signal for structure-based generative design, changing where generated substituents make contact rather than only how strongly they are predicted to bind. The gains on aggregate metrics are modest — a 0.08 kcal/mol shift in mean Vina score and under two points of high-affinity rate — so the per-target and per-interaction breakdowns carry more of the argument than the headline numbers. The evaluation is entirely computational, on a 49-pocket test set with docking-based scoring and no experimental validation, and the comparison is against a single baseline, the authors' own DiffDec. The work is an arXiv preprint and has not been peer reviewed. No DiffDecip code or checkpoint has been released; only the base DiffDec implementation is public, under an MIT license.