This study employed Gradient, an autonomous, AI-driven research-and-optimization framework, to systematically explore and refine the computational methodology. Gradient functions by defining a scientific objective and a corresponding evaluation metric, then iteratively proposing, implementing, and executing diverse method variants in parallel. This approach shifts the paradigm from hand-tuned pipeline development to automated search space exploration, concentrating computational resources on strategies that demonstrate statistically significant improvements in performance.
Methodological optimization was achieved using Seqera Gradient.
The primary objective was the prediction of the bound pose of small-molecule ligands within the ligand-binding domain (LBD) of the Pregnane X Receptor (PXR). The PXR LBD is characterized by a high degree of conformational flexibility and promiscuity, which enables the accommodation of diverse chemical structures. Given that the protein fold is well-defined, the central challenge involves accurate ligand-pose prediction within an adaptable binding pocket rather than protein fold determination.
We utilized modern AlphaFold3-class co-folding models, which jointly predict protein-ligand complexes directly from sequence and chemical input. Several open co-folding models were evaluated in this role, and the selection and sampling of these generative models were treated as dynamic variables within the optimization search, rather than fixed parameters.
The evaluation strategy was designed to provide a robust, leakage-aware internal signal for the automated search. We constructed a benchmark calculating true LDDT-PLI using OpenStructure against a curated panel of nuclear-receptor holo reference structures with known crystallographic ligand poses. This benchmark served as the primary optimization objective, with a small number of leaderboard submissions used to calibrate it so that internal gains could be read in terms of expected leaderboard movement. To ensure generalizability, we implemented strict leakage controls, excluding:
- Reference structures overlapping with the challenge holdout set.
- Deposition siblings of holdout entries.
- Ligand chemotypes exhibiting near-duplicate properties.
We systematically evaluated multiple method families to enhance ligand-pose accuracy, utilizing the leakage-controlled metric to assess efficacy:
- Base pose generators (co-folding models): Boltz-2 (challenge-class baseline, internal overall LDDT-PLI ≈ 0.60), Protenix and Chai-1 (both AF3-class, neither surpassed the baseline), IntelliFold-2 (an independent AF3-class model with its own architecture and training distribution — the strongest generator found, internal overall 0.637, leaderboard 0.512), and OpenFold3 (see fine-tuning below).
- Generation and Sampling Diversification: Broadening and re-ranking candidate pose pools via random-seed/diffusion-sample sweeps, MSA source and depth, holo structural templates, and reproducibility-consensus pool diversification. All were flat — pose accuracy proved generation-limited rather than sample-limited.
- Diffusion Guidance: Biasing generative sampling toward physically and geometrically plausible configurations. Native Feynman–Kac / SMC steering with physical-validity potentials gave the best result obtained from the Boltz-2 base (internal 0.617, leaderboard 0.502); test-time guided diffusion and reward-guided steering were clean negatives.
- Pose Refinement: Utilizing physics-based and machine-learning approaches for local energy minimization. Machine-learned force-field relaxation was flat; classical minimization degraded the score by breaking the scoring boundary; restrained minimization gave a marginal gain (≈ +0.01).
- Docking-Co-folding Hybrids: Regenerating ligand poses via docking within the pocket predicted by co-folding architectures. Fixed-receptor docking underperformed co-folding outright; the induced-fit hybrid (redocking into the co-folder's predicted pocket, selecting by decorrelated physics scores) showed real headroom on hard PXR references but was bottlenecked by the lack of a leakage-free pose selector.
- Advanced Pose Selection: Employing discriminative scoring signals beyond model confidence for ensemble-based pose selection. Learned critic re-ranking, validity filtering plus CNN selection, cross-engine routing, and analog-pose transplant were all flat to marginal; physics interaction-energy selection was still under evaluation at the time of writing.
- Receptor Conditioning: Pocket/receptor restraints plateaued, with the PXR-specific sub-metric falling below the unconstrained base.
- Target-Specific Fine-tuning: Fine-tuned an AlphaFold3-class co-folder (OpenFold3) on a small, strictly leakage-controlled set of nuclear-receptor and PXR co-crystal structures, specializing the model for this flexible pocket. This markedly improved the internal PXR-only sub-metric (≈ 0.41 → 0.57, and ≈ 0.60 with added best-of-N sampling) but did not improve the leaderboard score (0.447), as the leaderboard tracks the broader pose-accuracy distribution on which the un-fine-tuned IntelliFold-2 generator remains stronger.
- Different-Architecture Generator: Swapped the base co-folder for IntelliFold-2 to test whether a fresh generation distribution improves ligand-pose accuracy relative to the previous base — this proved to be the single largest source of improvement (see below).
A consistent finding was that, for this target, pose selection over a single generated pool is largely exhausted — model confidence, learned critics, and physics/interaction scores rank poses within run-to-run noise of one another. Leaderboard movement tracked the internal overall metric closely and near-linearly across three independently built submissions (leaderboard ≈ internal − 0.125), which is why the largest gains came from changing the generator itself rather than from re-processing a fixed pool of predictions.
The submitted configuration was IntelliFold-2, sampled with 3 random seeds × 5 diffusion samples per seed (a 15-pose pool) using a single pre-computed MSA and no structural templates, with the final pose selected by the model's own ranking score. This achieved an internal overall LDDT-PLI of 0.637 and a leaderboard LDDT-PLI of 0.512 — our strongest submission, and a faithful translation of the internal gain. A single-seed variant reaches essentially the same accuracy (0.632) at roughly one-third of the compute and is the preferred setting when folding at scale.