Source-linked AI summary
AtomSteer: Test-Time Constraint Steering for Atomistic Generative Models
Pengfei Cai, Katharina S. C. Jäger, Utkarsh Utkarsh, Nofit Segal, Bowen Deng, Akshay Subramanian, Rafael Gomez-Bombarelli
TL;DR
Atomistic generative models can produce structures that violate chemical and crystallographic requirements, while existing approaches lack a common test-time interface for heterogeneous constraints. AtomSteer steers pretrained models with bounded, structure-dependent projections during sampling. Across crystal, conformer, and protein–ligand tasks, it reports improvements in symmetry, stereochemistry, validity, and related outcomes.
Problem
Atomistic generative models can violate chemical and crystallographic requirements, and a common test-time method for heterogeneous constraints is missing.
Source: §1 · p. 1Method
AtomSteer pairs atomistic-graph constraint residuals with bounded, structure-dependent projections of pretrained models’ endpoint predictions, without retraining or architectural changes.
Source: §1 · p. 2Results
Across tasks, AtomSteer improves crystallographic symmetry and motif realization, conformer stereochemistry and ensemble properties, and protein–ligand geometry, validity, and joint success over pretrained models.
Source: §1 · p. 2, §5 · p. 10Takeaways & Limitations
AtomSteer provides a common test-time interface in which new structural requirements add constraint functions rather than requiring model retraining.
Source: §5 · p. 10Some cases show AtomSteer displacing a ligand without improving its stereochemistry.
Source: §4.3 · p. 10
Abstract
from arXiv · showhide
Diffusion and flow matching models offer scalable approaches to generating crystalline materials, molecular conformers, and protein-ligand complexes. However, satisfying crystallographic symmetry, structural geometry, and accurate stereochemistry remains challenging. Satisfying these requirements calls for corrections that preserve the geometry already captured by the pretrained model. In this work, we propose AtomSteer, a test-time inference framework for steering pretrained atomistic generative models with structural constraints. During sampling, AtomSteer corrects intermediate structure predictions with bounded local projections, graph-rigid stereochemical transformations, and projections over periodic symmetries, accepting updates that improve constraint satisfaction while limiting changes to the rest of the structure. We demonstrate AtomSteer across three pretrained backbones, FlowMM, ET-Flow, and Boltz-2, spanning de novo crystal generation and crystal structure prediction, molecular conformer generation, and protein-ligand complex structure prediction. AtomSteer improves space-group distribution coverage in de novo crystal generation and improves target space-group recovery in crystal structure prediction. AtomSteer improves molecular validity and E/Z stereochemical accuracy while retaining conformer ensemble fidelity. For protein-ligand complexes, AtomSteer reduces ligand E/Z errors relative to experimental reference structures while achieving competitive docking success and PoseBusters validity. Our method provides a unified framework for adding chemical and crystallographic control to existing atomistic generative models without retraining or architectural changes.
1 INTRODUCTION
AtomSteer addresses structural violations in atomistic generative models with bounded, test-time projections that preserve the pretrained model’s sampling process. Across crystals, conformers, and protein–ligand complexes, it reports improved symmetry, stereochemistry, and validity outcomes.
1 INTRODUCTION
- Pretrained atomistic models can violate chemical and crystallographic requirements, including ligand chemistry or chirality errors and oversampling low-symmetry crystals such as P1.The requirements matter because molecular graphs specify local geometry and stereochemistry, pose validity is needed for protein–ligand complexes, and space-group symmetries influence material properties.Source: §1 · p. 1
- Existing test-time methods lack a common interface for heterogeneous constraints whose measurements and appropriate corrective motions differ.For example, an E/Z error concerns four atoms but calls for rotating a substituent; symmetry concerns equivalent atoms and calls for moving their orbit together.Source: §1 · pp. 1–2
- AtomSteer pairs structural residuals on atomistic graphs with bounded, structure-dependent projections, steering pretrained models without retraining or architectural changes.Its projection families cover local geometry and stereochemistry, rigid fragments and bodies, and crystal symmetry and motifs.Source: §1 · p. 2
Crystallographic symmetry and motif control
- AtomSteer broadens generated space-group support and raises crystal-structure-prediction joint success@1 from 45.18% to 52.54%.The joint measure requires both a structure match and the correct space group; requested Archimedean lattice constraints also achieve a higher joint rate of satisfaction and SCIGEN prescreening.Source: §1 · p. 2
Conformer validity and ensemble quality
- On GEOM-Drugs, AtomSteer reduces median E/Z distribution error by 75.9% and GFN2-xTB ensemble-property errors by 10–36% versus the better ET-Flow vanilla sampler for each property.It also improves conformer precision and PoseBusters validity.Source: §1 · p. 2
Protein–ligand validity and stereochemical accuracy
- On StereoBench, AtomSteer achieves the highest PoseBusters validity and raises joint valid-docking-and-stereochemistry success from 36.5% to 45.2% versus Boltz-steering.Source: §1 · p. 2
2 RELATED WORK
Prior approaches incorporate constraints through model design, training, or inference-time guidance and projection. AtomSteer instead applies constraint-specific projections to pretrained models’ endpoint predictions during sampling.
Inference-time steering
- Inference-time steering includes trajectory reweighting, coordinate guidance, hybrid methods, and hard projections; AtomSteer projects endpoint predictions during sampling and returns accepted corrections through native updates without fine-tuning.The Boltz-steering baseline combines seven flat-bottom potentials with Feynman–Kac resampling; Chen et al. (2026) fine-tune the denoiser through end-of-denoising projection.Source: §2 · p. 3
Symmetry and constraints in crystal generation
- Crystal methods encode symmetry or chemistry in representations, conditioning, training, or inference; AtomSteer instead infers compatible symmetry and imposes motifs through test-time endpoint projection.Unlike the described alternatives, it does not require a target structure, supplied space group, or energy model.Source: §2 · p. 3
Molecular geometry and stereochemistry
- Molecular generative models encode geometry through graphs, fragments, torsions, objectives, or training data; AtomSteer projects graph-defined requirements onto pretrained endpoint predictions during inference.Source: §2 · p. 3
3 ATOMSTEER: TEST-TIME CONSTRAINT PROJECTION
AtomSteer represents each structural requirement as a residual on an atomistic graph and pairs it with a projection that moves relevant coordinates together. It intervenes on predicted endpoints during sampling, bounding and checking corrections before the native sampler continues.
3.1 STRUCTURAL RESIDUALS AND STRUCTURE-DEPENDENT PROJECTIONS
- Atomistic graphs group the atoms or crystal sites involved in a constraint and carry labels such as bond orders, stereochemistry, and fixed species; the generated configuration can include coordinates, lattice parameters, or atom types.Residuals measure violations, while the constraint set contains configurations where the residual is zero.Source: §3.1 · pp. 3–4
- Structure-dependent projections reduce violations by moving appropriate coordinates together and seek a satisfying structure with minimum displacement within a movement budget.Damped numerical corrections use a residual–displacement objective; λ_c controls the displacement penalty, and setting it to zero leaves displacement limited only by the movement budget.Source: §3.1 · p. 4
- When no space group is supplied, AtomSteer selects a compatible group from the endpoint prediction by promoting approximately satisfied groups to the most symmetric feasible parent, then projects positions and lattice metrics.Atomic types remain fixed during this projection.Source: §3.1 · p. 4
3.2 ENDPOINT PROJECTION DURING SAMPLING
- The three backbones provide endpoint predictions in distinct forms: FlowMM and ET-Flow extrapolate learned flow velocities, while Boltz-2 denoises noisy coordinates toward a clean estimate.FlowMM also applies representation operations, including periodic wrapping.Source: §3.2 · p. 4
- At scheduled steps, the projected endpoint replaces the original in the native update; bounded corrections that fail acceptance checks are discarded, preserving the pretrained model’s update for those coordinates.Table 1 groups structural constraints by projection family, and a single requirement can combine families, as with E/Z fragment operations followed by Cartesian ChemGN.Source: §3.2 · p. 4, Table 1 · p. 5
4 RESULTS
Across FlowMM, ET-Flow, and Boltz-2, AtomSteer is tested for constraint satisfaction and downstream outcomes. The reported results include improved crystal symmetry, conformer quality, and ligand stereochemistry.
4.1 ATOMSTEER IMPROVES CRYSTALLOGRAPHIC SYMMETRY IN FLOWMM
- 71.47%→77.91% symmetry success; AtomSteer broadens generated space-group support and lowers dSG toward MP-20, with clearer gains in larger cells.The symmetry gain persists after full-cell UMA relaxation.Source: Table 2 · p. 5, Figure 2 · p. 6, §4.1 · p. 6
- Joint@1 rises from 45.18% to 52.54% for composition-only AtomSteer with MR@1 essentially unchanged; lattice-informed AtomSteer reaches 66.33% MR@1 and 54.95% Joint@1.The lattice-informed variant receives a reference-derived lattice metric template and integer setting.Source: §4.1 · pp. 6–7, Table 3 · p. 7
- AtomSteer realizes requested Archimedean lattices in 96.6–99.2% of samples, on par with SCIGEN, and has the highest joint rate for all three lattices.After relaxation, the requested triangular lattice is detected in 86.4% of AtomSteer samples versus 31.5% for SCIGEN.Source: Table 4 · p. 7, §4.1 · pp. 7–8
4.2 ATOMSTEER IMPROVES CONFORMER VALIDITY, QUALITY, AND STEREOCHEMISTRY IN ET-FLOW
- AtomSteer lowers median E/Z distribution error by 75.9%, raises PoseBusters validity to 96.93%, and improves COV-P and AMR-P while COV-R remains 79.01% versus 79.93%.The comparison is against the stronger ET-Flow sampler, ET-Flow-SS.Source: Table 5 · p. 8, §4.2 · p. 8
- After GFN2-xTB relaxation, AtomSteer has the lowest median error on all four ensemble properties, 10–36% below the better ET-Flow sampler for each property.The properties are mean energy, dipole moment, HOMO–LUMO gap, and minimum energy.Source: Table 5 · p. 8, §4.2 · p. 8
4.3 ATOMSTEER IMPROVES E/Z GEOMETRY AND PHYSICAL VALIDITY IN BOLTZ-2
- On StereoBench, AtomSteer raises the within-10° fraction to 79.1%, 19.7 points above Boltz-steering, and lowers median E/Z error from 6.13° to 2.61°.The E/Z correction also holds on the E/Z subset of PoseBusters.Source: Table 6 · p. 9, Figure 3 · p. 9, §4.3 · p. 9
- Across both protein–ligand benchmarks, AtomSteer leads evaluated methods in PoseBusters validity, docking success, Success@1, and Success@5; on StereoBench, joint Success@1+stereo rises from 36.5% to 45.2%.R/S accuracy is at parity with Boltz-steering, and protein structure quality is mostly indistinguishable across methods.Source: Table 6 · p. 9, §4.3 · p. 10
- Some examples show AtomSteer displacing the ligand without a stereochemical gain.The paper also reports corrected and unchanged cases.Source: §4.3 · p. 10
5 CONCLUSION
AtomSteer provides a shared test-time interface for steering pretrained atomistic generative models with structural constraints. Its conclusion spans crystallographic, molecular, and protein–ligand properties, while identifying endpoint prediction and residual-plus-projection formulations as current requirements.
5 CONCLUSION
- AtomSteer improves crystallographic symmetry and motif realization, conformer stereochemistry and ensemble properties, and protein–ligand E/Z geometry, validity, and joint success over pretrained models.Each requirement is expressed as a structural residual paired with a structure-dependent projection; the framework currently requires a predicted endpoint or useful estimate and a corresponding projection.Source: §5 · p. 10
A.1 SAMPLING LOOP AND BACKBONE SCHEDULES
- At each sampling step, AtomSteer predicts an endpoint, applies scheduled bounded projections with acceptance checks, then resumes the backbone’s native sampler from the corrected structure.An unchanged prediction or rejected projection leaves the native sampler’s update intact.Source: §A.1 · p. 16
- Projection schedules vary by task and backbone; the retained samplers use 1,000 FlowMM steps, 50 ET-Flow steps, and 200 Boltz-2 diffusion steps.Table 7 specifies active projection families, their order, and when they act; ranges ending at 100% include the final update.Source: §A.1 · p. 16, Table 7 · p. 17
A.2 MOLECULAR CONSTRAINTS AND PROJECTIONS
- Molecular constraints are derived from the input graph, and endpoint residuals measure violations that projections reduce.ET-Flow and Boltz-2 share these geometric constraints and specified stereochemistry.Source: §A.2 · p. 17
A.2.1 BOND GEOMETRY, STERIC SEPARATION, AND PLANARITY
- Bond-length residuals enforce graph-derived distance intervals, while bond-angle constraints use graph-derived bounds on distances between atoms separated by two bonds.Reducing these 1–3 distance violations together with bond-length residuals constrains bond angles.Source: §A.2.1 · pp. 17–18
- Intramolecular clash residuals penalize eligible nonbonded pairs below a graph-derived minimum separation, while planarity residuals limit atom distances from designated ring or double-bond planes.Protein–ligand and intercomponent exclusions instead use atom-type radii and the surrounding protein.Source: §A.2.1 · p. 18
A.2.2 DOUBLE BOND STEREOCHEMISTRY AND TETRAHEDRAL CHIRALITY
- E/Z residuals measure wrapped angular deviation from the input-specified target, and tetrahedral residuals measure violations of the graph-derived signed-angle margin.The E/Z target is π for E and 0 for Z; reflection reverses oriented volume, which identifies tetrahedral handedness.Source: §A.2.2 · p. 18
A.2.3 STRUCTURE-DEPENDENT MOLECULAR PROJECTIONS
- Molecular projections combine local ChemGN updates with rigid-fragment or rigid-body moves that preserve internal geometry; global parity selection chooses between a conformer and its mirror image.E/Z fragment flips rotate eligible connected fragments while preserving their internal distances, followed by bounded refinement and local E/Z ChemGN corrections.Source: §A.2.3 · pp. 18–19
A.3 COMPARISON WITH RELATED METHODS
The crystalline-materials sections describe FlowMM integration, symmetry and motif projections, bounded correction checks, and evaluation protocols. The supplied passages outline these methods and protocols but do not report the corresponding benchmark outcomes.
A.3 COMPARISON WITH RELATED METHODS
- The related-methods comparison table records whether representative approaches demonstrate specified constrained-generation capabilities.Its checkmarks distinguish demonstrated, partial or variant-dependent, and un demonstrated capabilities.Source: Table 8 · p. 20, §A.3 · p. 20
B.1 FLOWMM INTEGRATION AND CRYSTAL PROJECTIONS
- FlowMM supplies the pretrained vector field and endpoint prediction used for de novo, crystal-structure-prediction, and motif-constrained tasks.Sampling uses 1,000 Euler steps; the de novo checkpoint serves de novo and motif tasks, while the CSP checkpoint serves CSP.Source: §B.1 · p. 20
B.1.1 CRYSTAL SYMMETRY PROJECTIONS
- Crystal symmetry projection jointly corrects symmetry-related fractional coordinates and a compatible lattice metric.Hierarchy-aware selection infers realizations from the endpoint prediction; composition-only CSP infers the metric template and orbit realization, while lattice-informed CSP receives a reference-derived template.Source: §B.1.1 · pp. 20–22
B.1.2 MOTIF-CONSTRAINED PROJECTIONS
- Motif projections enforce requested species, periodic lattice patterns, and nearest-neighbor spacing while retaining unconstrained cell parameters.The motif task covers triangular, honeycomb, and Kagome patterns; the prescribed spacing sets the basal lengths and angle.Source: §B.1.2 · p. 22
B.1.3 PROJECTION SCHEDULE AND ACCEPTANCE CHECKS
- Crystal corrections are rejected when cell or contact checks fail, and affected coordinates retain the native sampler update.Checks include finite, sufficiently large, well-conditioned cells and lattice-parameter reconstruction within numerical tolerance.Source: §B.1.3 · p. 23
B.1.4 FLOW CONTINUATION
- After projection, FlowMM continues from the corrected endpoint, preserving the native update for entries the projection leaves unchanged.Motif-constrained generation instead interpolates between initial noise and the projected endpoint.Source: §B.1.4 · p. 23
B.2.1 MP-20 DE NOVO GENERATION
- De novo symmetry evaluation uses 10,000 paired samples and compares detected space-group distributions against MP-20.The symmetry success rate excludes P1 and P-1, while distribution distance uses detected structures only; geometric validity is assessed before relaxation.Source: §B.2.1 · pp. 23–24
B.2.2 MP-20 CRYSTAL STRUCTURE PREDICTION
- CSP evaluation uses 9,046 targets with 20 samples per method, and joint success requires both reference-structure matching and space-group recovery.FlowMM and its AtomSteer variants share saved initial noise for each target.Source: §B.2.2 · p. 24
B.2.3 MOTIF-CONSTRAINED GENERATION
The motif-constrained evaluation tests requested Archimedean lattices under sampled conditions and measures constraint satisfaction, prescreening, and relaxation outcomes.
B.2.3 MOTIF-CONSTRAINED GENERATION
- The evaluation draws 1,000 conditions for each of Kagome, honeycomb, and triangular lattices using SCIGEN’s condition-sampling procedure.Elements are sampled uniformly from ten specified elements, with atom counts above the number of designated lattice sites and no more than 20.Source: §B.2.3 · p. 24
- A generated crystal satisfies an Archimedean-lattice condition only if it contains the requested element and lattice at the prescribed spacing and is geometrically valid.Detection checks periodic same-species subsets, positional RMS, in-plane lengths, and the 120° angle.Source: §B.2.3 · pp. 24–25
- Joint success counts samples that satisfy the lattice constraint and pass all four stages of SCIGEN prescreening.The evaluation also reports prescreening survival and the fraction of planned conditions whose relaxation yields a valid structure containing the requested shape.Source: §B.2.3 · p. 25
B.3 ADDITIONAL RESULTS FOR CRYSTALLINE MATERIALS GENERATION
The additional crystalline-materials results report force and relaxation outcomes, alongside figures for size-stratified space-group recovery and motif-constrained examples. The included molecular passages specify ET-Flow steering and GEOM-Drugs evaluation procedures.
B.3 ADDITIONAL RESULTS FOR CRYSTALLINE MATERIALS GENERATION
- 82.16% of AtomSteer structures versus 80.48% of vanilla FlowMM structures converge under full-cell relaxation; AtomSteer lowers maximum atomic force in 74.5% of paired samples.The comparison uses 10,000 paired samples.Source: §B.3 · p. 25
B.3.1 CRYSTAL STRUCTURE PREDICTION BY SIZE
- Figure 4 groups target SG@1 recovery for composition-only AtomSteer by generated unit-cell atom count across 9,046 MP-20 targets.The caption specifies the grouping and evaluation scope, but gives no values for the groups.Source: Figure 4 · p. 25
B.3.2 MOTIF-CONSTRAINED GENERATION EXAMPLES
- Figure 5 presents motif-constrained examples by target element, nearest-neighbor distance, and atom count, comparing requested, vanilla FlowMM, AtomSteer, and relaxed AtomSteer structures.Structures share tilt and length scale within each row.Source: Figure 5 · p. 26
C.1 ET-FLOW INTEGRATION AND MOLECULAR PROJECTIONS
- For ET-Flow conformer generation, AtomSteer uses 50 Euler steps and replaces post hoc chirality correction with global parity selection.The released ET-Flow and ET-Flow-SS baselines use deterministic and stochastic 50-step samplers, respectively.Source: §C.1 · p. 27
C.2.1 GEOM-DRUGS TEST SET AND CONFORMER BUDGETS
- The GEOM-Drugs evaluation excludes two disconnected molecules, leaving 998 molecules and 103,966 reference conformers.Each method generates 207,932 conformers on these matched ensembles.Source: §C.2.1 · p. 28
C.2.2 CONFORMER ACCURACY, VALIDITY, AND STEREOCHEMISTRY
- Conformer evaluation measures reference-to-generated coverage and RMSD, PoseBusters validity, and E/Z distribution error against held-out GEOM distributions.Coverage uses a 0.75 Å threshold; E/Z distribution error is summarized across 153 eligible molecules.Source: §C.2.2 · p. 28
C.2.3 GFN2-XTB EVALUATION OF ENSEMBLE PROPERTIES
The GFN2-xTB evaluation compares ensemble-property estimates with GEOM reference ensembles, while a separate table reports conformer coverage and matching statistics.
C.2.3 GFN2-XTB EVALUATION OF ENSEMBLE PROPERTIES
- GFN2-xTB evaluation measures median absolute errors for energy, dipole moment, HOMO–LUMO gap, and minimum energy against GEOM reference ensembles.It uses 100 randomly selected molecules, with generated conformers optimized before single-point calculations.Source: §C.2.3 · p. 28
C.3.1 CONFORMER COVERAGE AND MATCHING STATISTICS
- Table 12 reports mean and median conformer coverage and matching metrics across 998 GEOM-Drugs molecules, using a 0.75 Å COV threshold.The caption identifies best and second-best results through bold and underlining, but supplies no metric values.Source: Table 12 · p. 29
C.3.2 QUALITATIVE CONFORMER RESULTS
The examples show AtomSteer correcting highlighted conformer constraints and producing valid conformers across an ensemble, with stereochemical checks distinct from PoseBusters validity.
C.3.2 QUALITATIVE CONFORMER RESULTS
- Each Figure 6 example pairs an ET-Flow constraint violation with a nearby PoseBusters-valid AtomSteer conformer satisfying the highlighted constraint.The highlighted constraint is an E/Z dihedral, nonbonded distance, or bond length.Source: Figure 6 · p. 29
- PoseBusters settings in Figure 6 do not check E/Z configuration, so a conformer can pass those checks despite a stereochemical violation.Source: Figure 6 · p. 29
- Figure 7 shows three distinct AtomSteer conformers per molecule, all PoseBusters-valid and within 0.25–0.75 Å of their nearest GEOM reference.The figure covers five GEOM-Drugs molecules.Source: Figure 7 · p. 30
D.1 BOLTZ-2 INTEGRATION AND COMPARISON METHODS
Boltz-2 AtomSteer applies graph-derived molecular and intermolecular constraints to endpoint predictions during native sampling, using bounded, ordered projections and shared comparison settings.
D.1 BOLTZ-2 INTEGRATION AND COMPARISON METHODS
- Boltz-2 generates all-atom complex coordinates, while AtomSteer uses input graph relations and isomeric-SMILES stereochemical labels without experimental coordinates or reference labels.The setup uses 200 denoising steps, five returned samples per target, and native Boltz confidence for top-1 selection.Source: §D.1 · p. 31
D.1.1 LIGAND AND INTERMOLECULAR CONSTRAINTS
- Ligand constraints cover padded bond and 1–3 distance bounds, ring and double-bond planarity, E/Z and R/S stereochemistry, and ligand–protein exclusion.For exclusion, Cartesian ChemGN moves the ligand with the protein fixed, while rigid-body SE(3) moves the ligand as one body and preserves internal geometry.Source: §D.1.1 · p. 31
- For independently moving protein components, AtomSteer reduces overlap while keeping the ligand-contacting receptor component fixed and preserving other eligible components’ internal coordinates.Source: §D.1.1 · p. 32
D.1.2 PROJECTION SCHEDULE AND TERMINAL PROJECTIONS
- At each denoising step, endpoint projections run in a fixed order: E/Z refinement, Cartesian ChemGN, rigid-body projections, then bond-tangent R/S correction.Corrections are bounded and undergo residual and cross-constraint acceptance checks; settings and activation windows are specified in Tables 13 and 7.Source: §D.1.2 · p. 32, Table 13 · p. 32
- After denoising, terminal ligand, protein-component, and R/S corrections precede confidence selection; comparisons share checkpoint, seed, inputs, and native confidence selection.The compared alternatives include unsteered Boltz-2, guidance-only and Boltz-steering variants, and methods from Ishitani & Moriwaki and Chen et al.Source: §D.1.2 · pp. 32–33
D.2 BENCHMARKS
The benchmarks assess protein–ligand stereochemistry and pose validity on filtered experimental complexes, with common-target subsets used for comparisons.
D.2 BENCHMARKS
- StereoBench contains 1,103 experimental protein–ligand complexes with at least one non-ring stereogenic double bond and stringent structure-quality and ligand filters.Deposited coordinates are reserved for post-generation evaluation; E/Z metrics use all targets, while R/S metrics use the 538 with evaluable stereocenters.Source: §D.2 · p. 33
- The PoseBusters benchmark starts from 308 complexes; excluding three GPU-memory failures and four without an identifiable reference pocket leaves 301 common targets.Source: §D.2 · p. 33
- PoseBusters Success@5 uses 300 targets with complete results for all five poses and every method; its E/Z and R/S subsets contain 21 and 177 targets.All entries were released before the Boltz-2 training cutoff.Source: §D.2 · p. 33
D.3 EVALUATION METRICS
This section defines pose, validity, and stereochemistry metrics and describes additional StereoBench breakdowns and qualitative pose comparisons; the supplied passages do not report table-cell winners or values.
D.3 EVALUATION METRICS
- PoseBusters validity uses its 27-check redock profile against each predicted complex, excluding the unaligned ligand RMSD check; docking accuracy is evaluated separately after pocket alignment.Water and cofactor checks pass vacuously when those components are absent.Source: §D.3 · p. 33
- Docking success requires symmetry-corrected ligand RMSD below 2 Å after pocket alignment, while Success@1 additionally requires top-1 PoseBusters validity and Success@5 is an oracle over five poses.Top-1 is selected by Boltz-2 confidence.Source: §D.3 · p. 33
- E/Z error is the wrapped absolute predicted-versus-experimental dihedral difference, and stereochemical success requires every evaluable bond or center to meet its criterion.E/Z success uses a 10° limit; R/S success requires matching reference assignments.Source: §D.3 · p. 33
- OpenStructure independently scores protein and interface quality as a preservation check, but contains no stereochemical term.Source: Table 14 · p. 34, §D.3 · p. 34
D.4.1 E/Z ACCURACY BY BOND COUNT AND THRESHOLD
- StereoBench reports E/Z accuracy by double-bond count and threshold, covering 1,608 evaluable bonds across 1,103 ligands.The threshold table evaluates the percentage of ligands whose evaluable double bonds all fall within the stated angular threshold.Source: §D.4.1 · p. 34, Table 15 · p. 34, Table 16 · p. 34
D.4.2 PERFORMANCE BY PDB RELEASE DATE
- The release-date analysis recomputes Table 6 metrics for 926 pre-cutoff and 177 on-or-after-cutoff targets, with corresponding R/S subsets of 451 and 87.Source: §D.4.2 · p. 35, Table 17 · p. 35
D.4.3 QUALITATIVE POSE COMPARISONS
- Qualitative pose panels compare top-confidence poses against experimental ligands and annotate validity, ligand RMSD, and E/Z error; examples include cases where Boltz-steering has lower RMSD than AtomSteer.In Figure 10, all methods’ E/Z errors are within 10°; the examples report 1.2 Å versus 11.5 Å for 7G4D and 1.1 Å versus 5.3 Å for 7G6A (Boltz-steering versus AtomSteer). Figure 9 describes examples with all methods at RMSD ≤0.6 Å and most poses PB-valid.Source: Figure 8 · p. 35, Figure 10 · p. 36, Figure 9 · p. 36