Brian Jimenez

Harmonic Torsional Diffusion for Protein-Ligand Flexible Docking 1. Harmony is a flexible docking diffusion framework that treats ligand and pocket side-chain torsions as circular variables (on a torus) rather than regressing angle updates with Euclidean heads, addressing a key geometry mismatch in many diffusion docking models. 2. Core idea: parameterize each rotatable bond’s torsional score field as the derivative of a learned Fourier (harmonic) potential on the circle. Instead of predicting a single angle delta, the network predicts harmonic coefficients (a_k, b_k) that define a periodic potential whose gradient yields the torsional score. 3. The noise-level dependence for torsions is not left for the network to learn implicitly. Harmony uses an analytic factor from VE diffusion on the torus: the heat semigroup damps Fourier modes by exp(-(1/2) k^2 σ(λ)^2), inducing frequency-aware behavior where high-frequency components disappear at high noise and re-emerge as denoising proceeds. 4. This construction makes periodicity explicit, aligns the score parameterization with the actual forward noising process (wrapped Gaussians = heat kernel on the circle), and provides an inductive bias consistent with rotameric side-chain motion and multimodal ligand conformers. 5. Task setup: pocket-based flexible docking with fixed protein backbone and flexible pocket side chains. The state is modeled on a product manifold z = (t, R, τ, χ) in R^3 × SO(3) × T^{m_l} × T^{m_s}, jointly denoising translation, rotation, ligand torsions, and side-chain torsions. 6. Architecture: a heterogeneous SE(3)-equivariant e3nn GNN over a multi-scale pocket graph (ligand atoms, receptor residues at Cα with ESM-2 embeddings, and receptor atoms). Harmony is designed as a drop-in replacement for torsion regression heads: the backbone can stay the same while torsions become harmonic. 7. PDBBind (ESMFold apo receptors) results: with top-1 selection from 10 samples by an internal confidence head, Harmony improves ligand pose accuracy (L-RMSD < 2 Å: 64.2% vs ~39–42% for recent flexible pocket docking baselines) and improves pocket all-atom reconstruction (AA-RMSD < 1 Å: 47.3% vs 41.7% for FlexDock). It also reports faster runtime in their setup. 8. Physical validity: on PoseBusters without relaxation/post-processing, Harmony improves both success rate and “PB-valid” rates relative to comparable pocket flexible diffusion baselines, with notable gains in checks related to internal steric validity, ligand strain, and protein–ligand separation. 9. Ablations support the mechanism: fewer Fourier modes work best (peak around K=4), removing the analytic frequency damping hurts performance, and applying the harmonic head to only ligand or only side chains degrades the other—suggesting coupled learning benefits when both torsion sets use the harmonic formulation. 10. Case studies (EBNA1 6NPP and KRAS G12D 6GJ8) illustrate behavior in different pocket regimes: maintaining local loop closure in a polar/broader site and improving pose recovery in a shallow, exposed site where shape constraints are weak. 💻Code: github.com/dunnolab/harmony 📜Paper: arxiv.org/abs/2608.20366 #ComputationalBiology #MolecularDocking #DiffusionModels #GenerativeAI #DrugDiscovery #ProteinLigand #GeometricDeepLearning #SE3Equivariance #ScoreBasedModels

A Novel Benchmark Dataset for Enzyme Function Prediction Reveals the Limitations of State-of-the-Art Models 1. The paper introduces EnzymARC, a benchmark specifically designed to test whether EC-number predictors can detect loss of catalytic competence, not just recognize an enzyme family “scaffold” by sequence similarity. 2. EnzymARC is built from Swiss-Prot enzymes filtered to high-confidence annotations: 218,504 sequences with both an EC number and a Rhea ID (length 100–1,000 aa). Of these, 79,505 also include experimentally annotated active-site residues, enabling structure-guided decoy construction. 3. Key idea: generate “putative non-functional” decoys by systematically disrupting catalytic machinery while keeping the overall sequence highly similar—creating challenging negatives that standard benchmarks largely lack. 4. Two decoy types are created. (i) Catalytic decoys: mutate each annotated catalytic residue using ESM-2 masked LM, sampling from the five lowest-probability amino acids at that position (median identity to parent: 99.6%). (ii) Radius decoys: structure-aware random mutagenesis within 5 Å, 10 Å, and 15 Å shells around the active-site center of mass using AFDB structures (median identities: 98.4%, 89.8%, 72.9%). 5. Three representative EC prediction paradigms are benchmarked: DIAMOND (homology transfer), CLEAN (contrastive learning over ESM-1b embeddings), and DeepEC (CNN-based hierarchical EC predictor with an explicit enzyme vs non-enzyme discriminator). 6. DIAMOND largely fails to reject inactive decoys: it transfers the original EC to ~94–95% of catalytic/5 Å/10 Å decoys, and still yields ~89% false positives at 15 Å. This quantifies how global similarity transfer cannot detect active-site destruction. 7. CLEAN shows a similar vulnerability at low perturbation: false positive rates exceed 90% for catalytic and 5 Å decoys, and remain ~80% at 10 Å. Only at 15 Å does it substantially drop (to ~30.8% predicting the original EC), implying the learned embedding space often behaves like an enhanced similarity metric rather than a detector of catalytic integrity. 8. Importantly, CLEAN misclassifies with high confidence: decoy embeddings often remain close (low Euclidean distance) to the predicted EC cluster centers even after catalytic-site disruption, suggesting the model’s confidence calibration does not reflect mechanistic viability. 9. DeepEC is comparatively more sensitive when perturbations become large (FPR ~19% at 10 Å and ~4.4% at 15 Å), consistent with the benefit of training that includes negative examples (non-enzymes). But it still struggles with subtle active-site edits (FPR ~83.7% on catalytic decoys), indicating that generic non-enzyme negatives are not enough to learn active-site failure modes. 10. Takeaway: modern EC predictors can be dominated by “phylogenetic shortcuts,” failing to distinguish catalytically incompetent variants from functional enzymes when sequence scaffolds remain intact. The authors argue that structure-aware negative examples (like active-site-disrupted decoys) should be integrated into both training and benchmarking to improve functional robustness—especially relevant for genome annotation and for generative enzyme design pipelines that use EC numbers as optimization targets. 💻Code: github.com/jsartori12/EnzymA… 📜Paper: biorxiv.org/content/10.64898… #ComputationalBiology #Bioinformatics #Enzymology #ProteinLanguageModels #MachineLearning #Benchmarking #GenomeAnnotation #ProteinEngineering #AlphaFold #SwissProt