Team: Pavan Behara
Task: Structure track — predict the bound pose of 184 blinded PXR ligands
Best leaderboard result: LDDT-PLI 0.5174 · BiSyRMSD 3.7947 Å · LDDT-LP 0.9105 · Coverage 1 · Rank 7 (2026-06-22)
The submission uses Boltz-2 v2.2.1 [1], a diffusion-based joint structure prediction model for biomolecular complexes (co-folding). Boltz-2 generates full protein–ligand complex structures end-to-end from a protein sequence and a ligand SMILES string, with no physics-based docking step. The architecture is a Transformer-based structure module that operates in 3D coordinate space via diffusion, conditioned on a pairwise representation produced by an Evoformer-style trunk. Boltz-2 extends the AlphaFold 2 / AlphaFold 3 design lineage [2] with an affinity prediction head not used in this structure-track submission.
Each ligand was co-folded jointly with the full PXR sequence (aa 1–434). 100 independent diffusion trajectories (--diffusion_samples 100) were generated per ligand as the baseline. The best sample was selected by Boltz-2 ligand_iptm (the per-sample interface predicted TM score for the ligand chain from confidence_*.json), which empirically outperforms ipSAE-ligand as a selector on this task (rank ~19 → rank 11). ipSAE-ligand [3] — an inter-chain PAE-based interface confidence score adapted from the Dunbrack 2025 ipSAE metric — was used as a secondary selector and for model comparison. ipSAE-ligand filters protein residues to those with at least one cross-chain PAE < 10 Å, recomputes the Yang-Skolnick d0 [4] from the effective interface size, and averages the TM-score-weighted PAE contribution over the filtered interface.
Physical validity of each candidate pose was assessed with PoseBusters [10] (dock mode) prior to selection. The top-10 samples by ligand_iptm were checked; the highest-scoring physically valid sample was submitted. If all top-10 failed PoseBusters, the rank-1 sample by ligand_iptm was used without filtering. A supplementary geometric bond check (geom_ok) was applied to exclude rare Boltz-2 outputs with unphysical covalent bond lengths (e.g., N–C > 1.58 Å, C–C > 1.70 Å, S–C > 1.85 Å) that cause downstream OST scoring failures. Thresholds were set empirically from the observed distribution across all 184 predictions; 7 structures with stretched C–C or S–C bonds (1.62–1.69 Å) that nonetheless passed PoseBusters were retained by relaxing the thresholds to 1.70 Å and 1.85 Å respectively.
No post-processing (force-field minimisation, MLIP refinement, or rigid-body clash relief) was applied. An oracle-mode benchmark on PXR crystal structures reported by discoverybytes [5] showed that AMBER14sb, YAMBER3, MMFF94s energy minimisation, and MD all degraded LDDT-PLI relative to raw Boltz-2 output, confirming that post-processing is counterproductive for this system.
Targeted oversampling. Ligands with boltz_liptm_val < 0.93 in the Jun-16 submission (13 ligands) were identified as low-confidence cases likely to benefit from additional samples. An additional 100 diffusion trajectories were generated for each (--diffusion_samples 100 on a separate cluster run), bringing their total pool to 200 samples (or 300 for those that had previously received a 100-sample confidence rerun). The best sample from the merged pool was selected using the same ligand_iptm + PoseBusters top-10 criterion. Five of the 13 targets improved their selected liptm by >0.01 (x00242-1: +0.033, x01401-1: +0.039, x02848-1: +0.035), bringing them above the 0.93 threshold. The extreme outliers (x01438-1, x00046-1) showed only marginal improvement (+0.008, +0.007), suggesting fundamentally difficult binding modes for those compounds.
SMILES correction for 7 compounds. Seven ligands whose canonical SMILES contained non-standard bond-order representations (SMILES_FIX_IDS) were re-run with corrected molecular graphs. Their corrected predictions replaced the base boltz results in the final pool. The most impactful correction was x03462-1 (N–C bond 1.607 Å in the original prediction causing OST LDDT-PLI scoring failure), for which the constrained boltz run with corrected SMILES produced a geom_ok-passing pose.
Boltz-2 is pretrained on the Protein Data Bank (PDB) [6], including protein–small molecule co-crystal structures. No additional pre-training or fine-tuning was performed for this submission.
External data used at inference time (not training):
| Data source | Purpose |
|---|---|
| UniRef90 / BFD / MGnify | Multiple Sequence Alignment (MSA) generated via Boltz-2 MSA server |
| PDB templates | Structural templates enabled (--use_potentials, template search via Boltz-2 default pipeline) |
No proprietary data were used.
No hyperparameter optimisation was performed. Standard Boltz-2 inference settings were used: 100 diffusion samples, PDB output format, affinity potential enabled (--use_potentials). The ipSAE-ligand cutoff of 10 Å (PAE filter) and the Yang-Skolnick d0 formula were taken directly from the published metric definition without tuning. The choice of ligand_iptm as primary selector over ipSAE-ligand was determined empirically by comparing leaderboard LDDT-PLI across two submissions.
This task is treated as a 3D generative problem rather than a descriptor-based regression. The only molecular representation provided to the model is the ligand SMILES string, standardised through the OpenFF toolkit [7,8] pipeline:
LargestFragmentChooser(salt/solvent strip)Molecule.from_smiles(allow_undefined_stereo=True)— OpenFF atom-typing with SMIRNOFF perceptionmol.to_smiles()— canonical SMILES from OpenFF
This canonicalisation is identical to the pipeline used in the activity track submission to ensure consistent representation across tracks. Standardised SMILES are written into Boltz-2 YAML input files; no 2D or 3D descriptors are computed as model inputs.
Boltz-2 YAML input format (per ligand):
version: 1
sequences:
- protein: {id: A, sequence: <PXR_sequence>}
- ligand: {id: B, smiles: <canonical_SMILES>}
properties:
- affinity: {binder: B}Output CIF/PDB files were post-processed to rename chain-B residues to LIG (required by submission format), handling cases where OpenFold3 CIF files contain compound-registry residue names exceeding the 3-character PDB limit.
| # | Date | Rank | LDDT-PLI | BiSyRMSD | LDDT-LP | Selector | Notes |
|---|---|---|---|---|---|---|---|
| 1 | Jun early | 19 | 0.4738 | 4.1553 Å | 0.9118 | ipSAE-ligand | 5 samples |
| 2 | Jun early | 18 | 0.4810 | 4.1207 Å | 0.9090 | ipSAE-ligand | 100 samples |
| 3 | Jun 13 | 11 | 0.5068 | 3.9157 Å | 0.9100 | ligand_iptm + PoseBusters | 100 samples |
| 4 | Jun 16 | 10 | 0.5095 | 3.8955 Å | 0.9101 | ligand_iptm + PoseBusters | +x03462 geom fix; +6 SMILES fixes |
| 5 | Jun 18 | 33 | 0.4595 | 4.2478 Å | — | ipSAE × strain | regression — strain factor corrupted selector |
| 6 | Jun 18 | 12 | 0.4983 | 3.9598 Å | — | mixed_liptm | boltz_c / boltz_rocs swapped in for 7 ligands; slight degradation |
| 7 | Jun 18 | 15 | 0.4939 | — | — | ligand_iptm + PoseBusters | regression vs Jun 16; cause unknown |
| 8 | Jun 18 (resubmit) | 10 | 0.5095 | 3.8955 Å | 0.9101 | ligand_iptm + PoseBusters | resubmit of Jun 16 ZIP restored rank 10 |
| 9 | Jun 19 | 7 | 0.5166 | 3.8238 Å | 0.9108 | ligand_iptm + PoseBusters | +geom_ok relax; +low-liptm 100-sample rerun |
| 10 | Jun 22 | 7 | 0.5174 | 3.7947 Å | 0.9105 | ligand_iptm + PoseBusters + per-ligand model swap | x01438-1 → boltz_rocs (liptm 0.861→0.941, ipsae 0.609→0.824) |
| 11 | Jun 26 | 46 | 0.4224 | 4.6057 Å | 0.9073 | pure pharm_tsc | regression — template-specific contacts inflate liptm but corrupt pose placement |
| 12 | Jun 26 | 23 | 0.4891 | 4.0686 Å | 0.9102 | mixed_liptm hybrid (87 pharm_tsc + 97 boltz_base, per-ligand max) | regression — pharm_tsc component hurt even where liptm appeared to win |
| 13 | Jun 27 | 15 | 0.5134 | 3.8141 Å | 0.9112 | hard8 swap (8 hard cases → pharm_tsc, 176 boltz_base) | regression — constraints hurt even genuinely hard cases with large liptm/ipsae gains |
Key observations:
- Sample count (5→100): +0.007 LDDT-PLI (submissions 1→2). Modest gain — selector ceiling was the primary bottleneck.
- Selector (ipSAE→ligand_iptm + PB): +0.026 LDDT-PLI (submissions 2→3), BiSyRMSD −0.19 Å, rank 18→11. Ligand_iptm is a substantially better proxy for pose accuracy than ipSAE-ligand on this system.
- PoseBusters filter is load-bearing: Removing PB (submission 7, rank 15) costs ~0.013 LDDT-PLI vs the PB-enabled equivalent (submission 4, rank 10). Always apply PB pre-filter on the top-10 liptm candidates.
- SMILES fixes + x03462 geom fix (submission 3→4): +0.0027 LDDT-PLI. The x03462-1 N–C bond at 1.607 Å caused OST to score the pose as zero; substituting the corrected-SMILES constrained run recovered it.
- Boltz base > boltz_c / boltz_rocs (submission 4 vs 6): swapping 7 ligands to constrained/ROCS models lost −0.0085 LDDT-PLI despite those models having higher raw liptm. Use pure boltz base predictions only.
- Strain selector catastrophic failure (submission 5, rank 33): multiplying liptm by a strain penalty actively selected worse poses. Do not use MMFF strain as a selection criterion.
- geom_ok threshold relaxation + low-liptm targeted rerun (submission 8→9): +0.0071 LDDT-PLI, BiSyRMSD −0.072 Å, rank 10→7. Relaxing C–C (1.62→1.70 Å) and S–C (1.82→1.85 Å) thresholds unlocked 7 high-confidence poses (liptm 0.97–0.98) previously excluded by over-strict geometry checks. Additional 100-sample reruns for 13 low-liptm ligands moved 5 of them above the 0.93 threshold.
- Per-ligand model swap for extreme outlier (submission 9→10): +0.0008 LDDT-PLI, BiSyRMSD −0.029 Å. Swapping only x01438-1 from boltz_base to boltz_rocs (the one compound where ROCS-constrained predictions dramatically outperform: liptm 0.861→0.941, ipsae 0.609→0.824) confirmed that targeted per-ligand model selection is beneficial. The prior blanket 7-compound swap (submission 6) was counterproductive; a single well-justified swap is additive.
- Template constraints are uniformly harmful regardless of scope (submissions 11–13): tested at three granularities — pure pharm_tsc (184 ligands, rank 46, −0.095), mixed liptm-winner hybrid (87 ligands, rank 23, −0.028), and targeted hard-case-only swap (8 ligands, rank 15, −0.004). Every constrained approach regresses from rank 7. The regression is monotonically proportional to the number of constrained ligands swapped in (~0.0003 LDDT-PLI per ligand on average). Critically, the hard8 swap improved mean liptm (0.9604→0.9626) and ipSAE for all 8 hard cases yet still regressed — confirming that liptm and ipSAE improvements under template constraints are not reliable indicators of actual pose quality. The −0.004 hard8 delta falls within the 92-of-184 sampling noise window (±0.011–0.022 at 95% CI) and may be statistical noise rather than a true regression. Violin plot comparison:
outputs/plots/submission_violin_3way.png. - LDDT-LP ≈ 0.91 is stable across all valid submissions, confirming Boltz-2 reliably recovers correct ligand geometry. The remaining gap to rank 1 (LDDT-PLI 0.5725, Apheris) is in binding-pose placement accuracy, not ligand internal geometry.
- Gap to rank 1: 0.5725 − 0.5174 = 0.055 LDDT-PLI as of Jun 22.
Published oracle-mode benchmarks on PXR crystal structures (best of 50 samples per ligand) show Boltz-2 achieves LDDT-PLI ≥ 0.85 for 15/61 structures and ≥ 0.75 for 35/61 [5], outperforming other co-folding methods at the high-confidence tail. With only 5 samples selected by ipSAE-ligand rather than oracle selection, the gap between theoretical and practical performance is expected to be significant. The oracle benchmark suggests that increasing to 50 samples and applying a reliable selection metric could raise LDDT-PLI substantially.
Eight ligands with boltz_base liptm < 0.93. pharm_tsc improves their confidence metrics (liptm, ipSAE) but this does not translate to better leaderboard LDDT-PLI based on the hard8 swap experiment.
| ID | base liptm | base ipSAE | pharm_tsc liptm | pharm_tsc ipSAE | Δliptm | ΔipSAE |
|---|---|---|---|---|---|---|
| x01438-1 | 0.861 | 0.609 | 0.953 | 0.870 | +0.092 | +0.261 |
| x00046-1 | 0.889 | 0.740 | 0.948 | 0.860 | +0.059 | +0.121 |
| x01131-1 | 0.904 | 0.754 | 0.963 | 0.879 | +0.059 | +0.126 |
| x01464-1 | 0.904 | 0.760 | 0.963 | 0.893 | +0.059 | +0.133 |
| x00543-1 | 0.920 | 0.812 | 0.959 | 0.887 | +0.040 | +0.075 |
| x00337-1 | 0.920 | 0.812 | 0.944 | 0.873 | +0.024 | +0.061 |
| x01502-1 | 0.927 | 0.838 | 0.947 | 0.880 | +0.019 | +0.042 |
| x00990-1 | 0.929 | 0.822 | 0.952 | 0.868 | +0.022 | +0.047 |
Low ipSAE co-occurring with low liptm suggests the protein–ligand interface is poorly predicted for these compounds, not just sampling noise. Template constraints lift confidence metrics but apparently steer into wrong binding modes. Remaining approaches: OF3 predictions (50 samples, already run) or Chai-1 as complementary model.
The current gap to rank 1 (Δ LDDT-PLI ≈ 0.056) could be addressed through:
Deeper oversampling / template constraints for hard cases— tested exhaustively (pharm, pharm_tsc, hard8 swap). Constraints inflate liptm without improving LDDT-PLI. More boltz samples unlikely to help hard cases.ROCS-constrained predictions as a second pool— tested (submissions 6, 12); blanket swap consistently regresses. Only validated single-compound swap (x01438-1) was beneficial.- Chai-1 as a complementary model — oracle benchmarks show Chai-1 wins on different structures than Boltz-2; adding to the per-ligand selection pool could improve hard-case coverage without hurting easy ones. Not yet tested.
- OF3 predictions — 50 samples per ligand already run on cluster (
inputs/of3_predictions/). Analysis notebook:notebooks/of3_analysis.ipynb. Per-ligand winner selection between boltz_base and OF3 may cover hard cases where Boltz-2 sampling fails. Pharmacophore-guided template selection— implemented and tested (submissions 11–13). Uniformly harmful. See §S2.
Structural analyses of the PXR ligand-binding domain across 64+ co-crystal structures reveal that PXR accommodates diverse chemotypes through a combination of conserved hydrogen-bond anchors and adaptable hydrophobic contacts. Two residues — Ser247 (full-length numbering; LBD query position q106) and His407 (full-length; q266) — appear consistently across agonist co-crystal structures as the primary hydrogen-bond donors/acceptors for the ligand A-ring or carbonyl oxygen. The remaining binding-pocket contacts (a 77-residue union set) are occupied selectively depending on ligand size and shape. This observation motivates per-ligand template selection: a challenge ligand whose pharmacophore best matches a given crystal template is likely to dock into the same sub-pocket and make the same key contacts, giving that template higher information value as a Boltz-2 structural prior than a randomly selected or SMILES-similar template.
Schuster and Langer (2005) [ref 22] quantified the essential PXR pharmacophore as five features: one hydrogen-bond donor (HBD), one hydrogen-bond acceptor (HBA), two aromatic/hydrophobic centres, and a lumped hydrophobic body. This five-component pharmacophore vector, combined with ECFP4 Tanimoto similarity (which captures ring systems and substituent patterns), provides a chemotype-aware template ranking that preserves both 3D-shape and 2D-SAR information. A pure 3D shape score (ROCS) was shown to produce inferior predictions on average (submission 6, rank 12 vs rank 10), likely because ROCS ignores hydrogen-bond geometry — the very feature that determines which sub-pocket the ligand occupies.
Two Boltz-2 runs are planned from notebooks/pxr_pharm_template.ipynb:
Run 1 — Pharmacophore template + universal pocket (boltz_yamls_pharm/)
Template selected per ligand by score = 0.5 × cos(pharm_vec) + 0.5 × ECFP4_Tanimoto after a top-20 ECFP4 pre-filter over the 64 PXR co-crystal structures. The selected template CIF is passed to Boltz-2 as a structural prior. Pocket contacts use the universal 77-residue set (union of all contacts across all 64 templates), identical to the boltz_yamls_rocs/ runs. This isolates the effect of template selection alone from contact specificity.
Run 2 — Pharmacophore template + template-specific contacts (boltz_yamls_pharm_tsc/)
Same pharmacophore template selection, but pocket contacts are template-specific: only the residues observed within 6 Å of the co-crystal ligand in the matched template structure (computed via biotite sequence alignment → query-numbered residues). Typical contact set size: 15–40 residues vs 77 universal. The hypothesis is that restricting constraints to the sub-pocket actually occupied by the pharmacophore-matched template narrows the diffusion search space more precisely, at the cost of potentially missing contacts that the challenge ligand exploits but its crystal template analog did not.
| Property | pharm | pharm_tsc |
|---|---|---|
| Template source | Top-1 by 0.5×cos(pharm)+0.5×ECFP4 |
Same |
| Pocket contacts | 77 residues (universal) | 15–40 residues (template-specific) |
| YAML dir | inputs/boltz_yamls_pharm/ |
inputs/boltz_yamls_pharm_tsc/ |
| SLURM script | slurm/boltz_predict_pharm.sh |
slurm/boltz_predict_pharm_tsc.sh |
| Diffusion samples | 100 | 100 |
The two runs will be compared against the existing boltz_liptm best (100 samples, default template search, 77-residue universal contacts). If pharmacophore template selection improves liptm for the 8 remaining hard cases (median liptm < 0.93), it will be incorporated into per-ligand pool selection alongside the existing boltz_base and boltz_rocs pools.
The author thanks discoverybytes for the detailed oracle benchmark of co-folding models on PXR crystal structures, which informed the choice of Boltz-2 and the no-refinement strategy. The OpenADMET team is gratefully acknowledged for providing the challenge infrastructure and tutorial pipeline.
[1] Wohlwend J, Corso G, Passaro S, et al. Boltz-1: Democratizing biomolecular interaction modeling. bioRxiv. 2024. doi:10.1101/2024.11.19.624167. (Boltz-2 v2.2.1 used in this submission.)
[2] Abramson J, Adler J, Dunger J, et al. Accurate structure prediction of biomolecular interactions with AlphaFold 3. Nature. 2024;630:493–500.
[3] Gao Z, Joseph NT, Soni N, Dunbrack RL. A ranking metric based on predicted aligned error for protein–protein and protein–nucleic-acid complexes. bioRxiv. 2025. (ipSAE metric; adapted here for protein–small molecule interfaces.)
[4] Yang H, Skolnick J. Specific interactions for ab initio folding of protein terminal regions with secondary structures. Proteins. 2004;55(3):519–529. (d0 formula underlying TM-score and ipSAE.)
[5] discoverybytes. Oracle-mode evaluation of co-folding models on PXR crystal structures, PXR Challenge community post. 2025.
[6] Berman HM, Westbrook J, Feng Z, et al. The Protein Data Bank. Nucleic Acids Res. 2000;28(1):235–242.
[7] Mobley DL, Bannan CC, Rizzi A, et al. Escaping atom types in force fields using direct chemical perception. J Chem Theory Comput. 2018;14(11):6076–6092.
[8] Wagner J, Thompson M, Mobley DL, et al. openforcefield/openff-toolkit: 0.18.0. Zenodo. 2025. https://doi.org/10.5281/zenodo.17525527.
[9] Björn Andreassen O, Mirdita M, et al. OpenFold3: Open-source implementation of AlphaFold3. v0.4.1. GitHub. 2025.
[10] Buttenschoen M, Morris GM, Deane CM. PoseBusters: AI-based docking methods fail to generate physically valid poses or cannot be correctly applied to real drug discovery tasks. Chem Sci. 2024;15(9):3130–3139.
Reference [3] defines ipSAE for protein–protein interfaces using cross-chain PAE matrices from structure prediction models. The implementation used in this submission adapts the metric to protein–small molecule complexes. This section derives both formulations and documents all differences.
Let chain A have
Step 1 — interface residue sets (both chains):
Step 2 — effective interface size and Yang–Skolnick d₀ [4]:
Step 3 — per-residue TM score, averaged over interface:
Difference 1 — Token granularity on the ligand side.
In Boltz-2 and OpenFold3, the ligand is tokenised one token per heavy atom, not one token per residue. The PAE cross-block is
Difference 2 — One-sided filtering for d₀.
Only the protein side is PAE-filtered; all ligand atoms are always included in the interface count:
Filtering the ligand side is omitted because for a near-native pose essentially every heavy atom is at the binding interface — PAE-filtering would drop most of them and collapse d₀ to 1.0 Å (maximally strict). Including all ligand atoms keeps d₀ in the ~2 Å range appropriate for protein–protein interfaces:
| Scenario |
|
d₀ |
|---|---|---|
| Protein–protein (25 + 25 interface residues) | 50 | 2.23 Å |
| ipSAE-ligand (25 good prot res + 30 lig atoms) | 55 | 2.44 Å |
| Ligand treated as 1 residue | 26 | 1.0 Å (clamped) |
Difference 3 — Normalization and averaging are protein-side only.
The inner sum normalises by the number of sub-threshold ligand atoms per protein residue; the outer average is over
| Property | Original ipSAE [3] | ipSAE-ligand |
|---|---|---|
| Partner token unit | residue | heavy atom |
| d₀ interface count | |A^int| + |B^int| (both filtered) | |A^good| + N^atoms_lig (protein filtered; all lig atoms) |
| PAE filtering on partner side | yes | no |
| Inner normalization | # sub-θ partner residues | # sub-θ ligand atoms |
| Outer average | over A^int | over A^good |
| Score range | [0, 1] | [0, 1] |
Difference 2 is structurally the most consequential: including all ligand atoms in d₀ keeps the scoring regime physically calibrated for small molecules. Difference 1 (atom-level tokenisation) is a consequence of how Boltz-2 and OpenFold3 represent ligands in their PAE matrices, not a design choice.
BiSyRMSD (Binding-Site symmetry-corrected RMSD)
Ligand RMSD computed after superimposing the binding-site protein atoms (rather than the full chain) and after resolving symmetry-equivalent atom orderings in the ligand. Used as the primary pose accuracy metric in this challenge.
Co-folding
Joint end-to-end prediction of a protein–ligand (or protein–protein) complex structure directly from sequence and SMILES, without a separate docking step. Boltz-2 and OpenFold3 are co-folding models.
Coverage
Fraction of challenge compounds for which a valid prediction was submitted. Coverage = 1 means all 184 compounds were predicted successfully.
d₀
Distance normalisation parameter from Yang and Skolnick [4] that calibrates the TM-score denominator to the size of the structure. Defined as
Diffusion samples (--diffusion_samples)
Number of independent stochastic reverse-diffusion trajectories run per input. Each trajectory produces a different candidate pose. The best is selected by ipSAE-ligand; oracle selection (best by ground-truth LDDT-PLI) sets the theoretical ceiling.
ipSAE (Interface PAE Score)
Confidence metric for predicted protein–protein interfaces defined by Gao et al. [3]. Uses PAE cross-terms between the two chains, filters to interface residues on both sides, and averages a TM-score-weighted contribution. See Supplementary S1 for the full derivation.
ipSAE-ligand
Adaptation of ipSAE [3] to protein–small molecule complexes used in this submission. Differences from the original: ligand tokens are per-atom (not per-residue); all ligand atoms contribute to d₀ (no PAE-filtering on the ligand side); averaging is protein-side only. See Supplementary S1 for the full derivation.
ipTM (Interface predicted Template Modelling score)
AlphaFold 3 / Boltz-2 output confidence score measuring inter-chain structural agreement. Defined as the TM-score analogue computed over inter-chain token pairs using the predicted distance error
LDDT (Local Distance Difference Test)
Per-atom score measuring the fraction of interatomic distances within a neighbourhood (2 / 4 / 8 / 16 Å shells) that agree between predicted and reference structures to within a tolerance. Ranges from 0 (no agreement) to 1 (perfect agreement). Robust to global superposition errors.
LDDT-LP (LDDT — Ligand Pocket)
LDDT computed over ligand–protein contact distances only; reflects the quality of ligand internal geometry and local protein–ligand contacts, not global pose placement. LDDT-LP ≈ 0.91 indicates correct bond lengths and angles.
LDDT-PLI (LDDT — Protein–Ligand Interface)
LDDT computed over all interatomic distances within 4 Å of the protein–ligand interface; the primary quality metric for pose accuracy in this challenge.
MLIP (Machine Learning Interatomic Potential)
Neural-network-based potential used for structure relaxation or molecular dynamics (e.g., YAMBER3, MACE). Oracle benchmarks show MLIP refinement degrades LDDT-PLI for Boltz-2 outputs on PXR.
MSA (Multiple Sequence Alignment)
Alignment of homologous sequences used to infer co-evolutionary information. Boltz-2 and OpenFold3 use MSAs derived from UniRef90 / BFD / MGnify to populate the Evoformer pairwise representation.
PAE (Predicted Alignment Error)
pae_*.npz; OpenFold3: *_confidences.json). Entry
PDB (Protein Data Bank)
Both the primary archive of experimentally determined macromolecular structures (rcsb.org) and the .pdb flat-file format used for coordinate exchange. Boltz-2 predictions are requested in PDB format (--output_format pdb) to match the challenge submission requirement.
PXR (Pregnane X Receptor)
Nuclear receptor (NR1I2) that acts as a master regulator of xenobiotic metabolism. PXR agonism induces CYP3A4 and other drug-metabolising enzymes, causing clinically significant drug–drug interactions. The binding pocket is large and malleable, accommodating structurally diverse ligands.
SMILES (Simplified Molecular Input Line Entry System)
Text-based representation of molecular structure. Ligand inputs to Boltz-2 and OpenFold3 are provided as canonical SMILES strings standardised via the OpenFF toolkit [7,8].
TM-score (Template Modelling score)
Length-normalised structural similarity score ranging from 0 to 1. Values > 0.5 generally indicate the same fold. The per-token TM formula
YAML
Human-readable configuration format (.yaml) used for Boltz-2 input files. Each file specifies the protein sequence (chain A), ligand SMILES (chain B), and optional constraints (templates, pocket residues, affinity properties).