Ross Irwin

Ensemble-Conditioned Molecular Design 1. The paper reframes molecular design from “find a single bioactive conformer” to “optimize the conformational ensemble”: both (i) ensemble modes (distinct low-energy states such as shapes, pharmacophore arrangements, or pocket-bound poses) and (ii) ensemble properties (aggregate scalars over the ensemble such as mean PSA or mean pairwise RMSD). 2. Core method: ensemble-conditioned guidance for 3D molecular generative models. Instead of training on paired multi-state data, the model is trained on single conditions and multiple conditions are composed at inference by combining the vector fields (continuous coordinates) and log-probabilities (discrete atoms/bonds) produced under each condition. 3. This enables a single trained model to express many design tasks by choosing inference-time compositions: target multiple modes, avoid specific modes (negative weights), mix modalities (shape + pharmacophore + pocket + property targets), and combine arbitrary numbers of conditions—without rebuilding datasets per task. 4. Mode conditions supported: (a) shape profiles represented as noisy coordinate point clouds with tunable fidelity (sigma controls how tightly the generated molecule must match), (b) pharmacophore point clouds with types and optional direction vectors, and (c) “soft” protein pocket conditioning (generate a ligand that fits a pocket without forcing a specific pose). 5. Property conditions supported (demonstrated): mean polar surface area (PSA) across conformers and mean pairwise RMSD as a flexibility proxy. These are provided as target scalar values and integrated via conditioning embeddings, allowing simultaneous control of a bound mode and whole-ensemble behavior. 6. A key technical issue is composing structural conditions from different reference frames. The paper introduces adaptive symmetry learning: encoders are trained to output either equivariant or invariant features depending on whether the input is randomly rotated during training, letting one condition define the generation frame while others are safely composed invariantly. 7. The generator uses joint continuous-discrete flow matching: coordinates are generated by continuous flow matching, while atom types and bond types are generated via discrete flow matching (CTMC). The guidance composition is designed to work consistently across both continuous and discrete outputs. 8. Flexible-size generation is added to pocket-conditioned 3D generation by padding to a maximum atom count (48 heavy atoms) and learning to emit padding tokens. Two training details are highlighted as important: placing pad coordinates at the molecule center-of-mass and aligning padded prior/data atoms with a permutation that minimizes squared distance. 9. Benchmarks introduced: (i) multi-mode conditioning with mode targeting and mode avoidance using compact vs extended conformer shapes; results show strong gains over a size-matched virtual screening baseline, with sigma trading off conditioning fidelity vs chemical novelty. (ii) ensemble property optimization while maintaining pocket + interaction pharmacophore constraints; PSA and RMSD respond monotonically to targets while interaction recovery and docking scores remain largely stable (directional rather than perfectly calibrated control). 10. Case studies: (a) dual-target design for AChE (donepezil) and MAO-B (safinamide) by composing two pharmacophore conditions; docking-based evaluation shows higher fraction of molecules scoring well in both pockets than single-target conditioning. (b) active-state-selective agonist design for A2A adenosine receptor by targeting the active pocket while avoiding the inactive pocket; adding the negative inactive-state condition increases active-state preference in docking (∆Vina state) while keeping low similarity to the known agonist. 📜Paper: arxiv.org/abs/2609.15077 #ComputationalChemistry #Cheminformatics #MolecularGeneration #DiffusionModels #FlowMatching #DrugDiscovery #GenerativeAI #StructureBasedDesign #MultiObjectiveOptimization #ConformerEnsembles

Boltzmann-Expected Molecular Design with Decoupled Annealing Flows 1. The paper reframes 3D molecular design around the key fact that many relevant 3D observables (e.g., free energies, shape descriptors) are expectations over a molecule’s Boltzmann conformational ensemble, not properties of a single generated conformer. 2. It introduces DECAF (Decoupled Annealing Flows), which factorizes the joint distribution over molecular graphs and 3D coordinates into two conditional flow models: a graph-conditioned coordinate flow p(x | G) that emulates the Boltzmann ensemble, and a coordinate-conditioned graph flow p(G | x) that proposes new graphs using 3D information. 3. DECAF alternates between these two flows in a simulated-annealing-like loop: sample an ensemble X from p(x | G), propose a new graph G’ via p(G | x) using coordinates from the ensemble, resample an ensemble for G’, and accept/reject the graph move with an annealing acceptance rule based on an ensemble-scored objective f(X). 4. A central design choice is that the scoring function is evaluated on ensemble statistics estimated from i.i.d.-style samples from the Boltzmann emulator p(x | G). This makes “ensemble-consistency” the optimization target, addressing a failure mode of property-guided 3D generators that can optimize for a single favorable conformer while the true ensemble violates the objective. 5. The approach is modular: changing the optimization objective does not require retraining the generative models. Any computable ensemble-level score (including moment-based objectives) can be plugged into the annealing loop. 6. On GEOM-Drugs, DECAF optimizes ensemble means of radius of gyration (Rg) and solvent-accessible surface area (SASA). The study shows that using larger ensembles during optimization (e.g., N=10 or N=50 vs N=1) is especially important for larger, more flexible drug-like molecules where the Boltzmann distribution is broader; single-conformer optimization degrades notably beyond ~50 atoms. 7. The paper evaluates optimization shifts using Cliff’s delta, reporting both local shifts along trajectories and global shifts across the dataset, oriented so that successful minimization/maximization corresponds to strong negative/positive shifts respectively. 8. DECAF extends naturally to multi-objective optimization over coupled geometric properties (Rg and SASA) using an augmented Tchebycheff scalarization, enabling trade-offs even when objectives are correlated and when their desired directions conflict (e.g., maximize Rg while minimizing SASA). 9. A distinctive capability is higher-moment “distribution shaping”: DECAF jointly optimizes variance and skewness of an ensemble property (demonstrated for Rg) to design molecules that are both flexible (high variance) and biased toward a prescribed conformational regime (via skewness control). The resulting conformational distributions are validated with all-atom MD simulations, not just the learned emulator. 📜Paper: arxiv.org/abs/2607.19519 #ComputationalBiology #MolecularDesign #GenerativeModels #FlowMatching #NormalizingFlows #DrugDiscovery #MolecularDynamics #Boltzmann #MultiObjectiveOptimization #MachineLearning

Confidence Is the Key: How Conformal Prediction Enhances the Generative Design of Permeable Peptides 1. The paper argues that RL-guided generative design can be misled by predictive models when generation drifts outside the model’s applicability domain, producing “high reward, high uncertainty” peptides—especially problematic for understudied cyclic peptides. 2. They integrate conformal prediction (CP) directly into the reinforcement learning (RL) scoring loop (PepINVENT-style peptide completion), so the agent is rewarded not just for predicted permeability, but for permeability predictions made with calibrated confidence at a user-chosen confidence level (here, 80%). 3. Core technical setup: a permeability classifier (XGBoost on ECFP features) is trained on CycPeptMPDB PAMPA data (6876 cyclic peptides; threshold LogPexp ≥ -6 as permeable). On top of this, they build an aggregated Mondrian inductive conformal predictor (ACP with 10 ICPs) outputting two p-values: P1 (permeable) and P0 (non-permeable). 4. Key conceptual point: CP’s two p-values encode “evidence for each class,” enabling four outcomes (Class 0, Class 1, Both, None). In RL, the target is not merely “high P(permeable)” but conformal efficiency: confidently permeable designs where P1 > 0.2 and P0 < 0.2 (at significance 0.2). 5. Baseline finding: optimizing raw model probability (standard practice) increases average predicted permeability (raw score rises ~0.51→0.87 over 350 epochs), but many “permeable” designs are not conformally confident—highlighting a mismatch between probability-based rewards and calibrated reliability. 6. They test multiple CP-based reward designs: maximize P1, maximize (1−P0), maximize (P1−P0), plus two discrete schemes: “harsh” (reward 1 only if both thresholds met) and “soft” (reward 1 if both met, 0.5 if one met, else 0). 7. Main methodological takeaway: single p-value optimization (P1 alone or 1−P0 alone) is learnable but does not reliably increase the number of confidently permeable peptides, because maximizing P1 does not ensure low P0 (and vice versa). The joint decision structure of Mondrian ICP matters. 8. Best-performing strategy: the CP “soft” scoring function converges fastest to the desired region (defined as reaching ~50% conformally efficient permeable predictions among valid molecules) and yields more reliable hits than raw-probability scoring when “hits” are defined as confident + within-domain. 9. Practical insight on generation dynamics: the soft reward reduces brittleness from sparse rewards (compared to harsh) and improves efficiency—fewer unique valid molecules may be generated overall, but a higher fraction meet the calibrated confidence criterion, meaning less wasted exploration in uncertain space. 10. Robustness check: performance depends on peptide length and training-data coverage. The CP-soft approach works well for lengths well represented in training (6, 7, 8, 10), but deteriorates for 9, 11, 12, effectively flagging when the predictor’s applicability domain is being exceeded—useful as a “stop relying on this objective” signal. 📜Paper: arxiv.org/abs/2605.05770 #ConformalPrediction #ReinforcementLearning #GenerativeModels #PeptideDesign #CyclicPeptides #UncertaintyQuantification #Cheminformatics #ComputationalBiology #DrugDiscovery #MachineLearning