Geometric Relevance Is Not Sufficient: Evaluating Focused Post-Training Supervision in Diffusion Docking

Michael Brown; Maryam Rahimimovassagh; Mehdi Yazdani-Jahromi; Sanjeda Sara Jennifer; Ivan Garibay

cs.LGNov 5, 2026

Geometric Relevance Is Not Sufficient: Evaluating Focused Post-Training Supervision in Diffusion Docking

Authors: Michael Brown, Maryam Rahimimovassagh, Mehdi Yazdani-Jahromi, Sanjeda Sara Jennifer, Ivan Garibay

Organizations: University of Central Florida, Orlando, FL, USA

Abstract

Diffusion-based molecular docking requires learning a corrective score field across a broad corruption distribution, yet individual pose errors can differ substantially in their structural consequences. This creates a post-training conflict between preserving broad score-field coverage and concentrating supervision on corrections that appear more relevant to docking geometry. Using DiffDock, a fixed-budget study compares broad continuation with torsional supervision focused on Cartesian consequence or conditional target consistency, and with a broad/focused mixture. Evaluation separates fixed score fidelity, generated-pose quality, best-of-10 sampling capacity, and confidence-selected Top-1 docking performance. At <5 Angstrom, broad continuation had the highest mean Top-1 success among trained policies, and focused or mixed supervision showed no consistent advantage under the tested budget. The starting model had better fixed score fidelity, whereas the evaluated broad-continuation model had higher mean Top-1 success with the same aggregate best-of-10 recovery. Focusing supervision on structurally consequential or more consistently specified corrections did not consistently improve returned-pose success over broad continuation.

Figures & tables

Policy	Fixed loss	Generated $<5$ Å	Best-of-10 $<5$ Å	Top-1 $<5$ Å
Starting model	1.6034	45.76	86.87	34.34
Broad intrinsic	1.6546	47.14	86.87	40.07
Broad Cartesian	1.6559	46.84	85.86	38.38
Count-matched intrinsic	1.6532	47.10	87.21	37.37
Count-matched Cartesian	1.6538	46.90	85.52	38.05
Consistent intrinsic	1.6519	46.84	85.86	38.38

Table 1: Main post-training outcomes. Docking columns are percentages; fixed score loss is lower-is-better. The starting model is one checkpoint; trained-policy values are three-seed means.

Figure 1: (a) Bad-pose rates for all states and the highest tail-energy decile; (b) crossed Cartesian-consequence and target-consistency assignments; and (c) direct diagnostic and analysis-only reference alignment with the score target.

Appendix figures & tables15 assets

Supplementary material from the paper’s appendix.

Appendix

Policy	Allocation	Torsional form or objective	Role in the comparison
Starting model	none	none	Unchanged reference model
Broad intrinsic	all torsions	$I$	Broad continuation
Broad Cartesian	all torsions	$\widetilde{G}$	Cartesian weighting without selection
Count-matched intrinsic	permuted subset	$C_{\pi}$	Generic sparse-allocation control
Count-matched Cartesian	permuted subset	$C_{\pi}\widetilde{G}C_{\pi}$	Sparse Cartesian control
Consistent intrinsic	consistency mask $C$	$C$	Target-consistency allocation

Appendix

Table 2: Definitions of the starting model and seven trained policies. The starting model is evaluated once; each trained policy is evaluated with three training seeds. The count-matched mask preserves the number of selected torsions within each graph while changing their assignments.

Slice	$n$	Spearman $\rho$	Baseline bad	High-tail bad	Risk difference
PDBBind train	1000	0.171	0.555	0.900	0.345
PDBBind validation	197	0.253	0.579	0.850	0.271
PDBBind test, no receptor overlap	48	0.164	0.521	0.800	0.279
MOAD train	1000	0.155	0.607	0.880	0.273

Appendix

Table 3: Definition and summary of the

k=1

tail-energy association across the PDBBind train, validation, and test slices and MOAD train slice. These are diagnostic summaries, not trained-policy comparisons.

Figure 2: Tail-energy association with difficult docking states in the 1,000-graph PDBBind training slice. Bars show pose-outcome rates across the full slice and within the highest 10% and 5% of tail energy. This diagnostic relationship does not establish a corrective training rule.

Signal use	PDBBind train	PDBBind val	PDBBind test	MOAD train
Standalone tail energy	0.171	0.253	0.164	0.155
Contextual fusion	0.238	0.293	0.127	0.233
Translation-gated fusion	$-0.252$	$-0.255$	$-0.318$	$-0.213$
Rotation-gated fusion	$-0.230$	$-0.230$	$-0.308$	$-0.187$

Appendix

Table 4: Representative Spearman associations between signal variants and final RMSD. The two noise-dependent gated variants reverse the positive association observed for standalone tail energy and contextual fusion across the four evaluated slices.

Direction choice	Component	$n$	Mean cosine	Positive fraction
Direct finite difference	Translation	1,574	0.147	59.6%
Direct finite difference	Rotation	1,548	$-0.035$	46.1%
Energy-ranked	Translation	1,574	0.118	59.8%
Energy-ranked	Rotation	1,548	$-0.048$	46.3%
Score-aligned axis reference	Translation	1,574	0.599	92.2%
Score-aligned axis reference	Rotation	1,548	0.496	84.7%

Appendix

Table 5: Alignment of tested direction choices with the corresponding score target. The diagnostic choices remain weakly aligned, whereas the score-aligned axis and free-sign rows use analysis-only information and serve as references rather than deployable corrective rules. The pure RMSD reference is defined separately.

Figure 3: Directional diagnostic and analysis-only reference comparison. Direct diagnostic directions are compared with the score-aligned axis reference used in the alignment summary. The analysis-only reference is unavailable to the docking model and is used to assess diagnostic-to-action agreement.

Gate cutoff	Reference score loss	Global cosine	Rotation-head cosine	Batches
2.0	Rotation component	0.026	$-0.064$	20
2.0	Full score loss	$-0.044$	$-0.064$	20
3.0	Rotation component	0.026	$-0.158$	20
3.0	Full score loss	$-0.009$	$-0.158$	20

Appendix

Table 6: Representative mean cosine alignments between a gated geometric objective and score losses over 20 valid batches. Near-zero global values and negative rotation-head values show weak or conflicting alignment; rows with no valid batches are omitted.

	Consistent row	Inconsistent row
High Cartesian consequence	349	279
Low Cartesian consequence	645	339

Appendix

Table 7: Crossed supervision priorities among 1,612 time-indexed qualification observations from 403 torsion-edge groups and four noise levels. Rows use the static

90^{\circ}

cross-pair screen and the score-target consistency threshold defined in the text.

Source	Processed	Multi-torsion	Metric states	Direction tests	Graph failures
PDBBind	192	173	1656	13248	0

Appendix

Table 8: Coverage and finite-difference checks for the local Cartesian metric on the PDBBind qualification set. These checks qualify the metric approximation and do not establish downstream training benefit.

Contrast	Generated $<5\,\text{\AA}$	Best-of-10 $<5\,\text{\AA}$	Top-1 $<5\,\text{\AA}$
Broad $-$ starting	$+1.38\ [0.17,\,2.56]$	$+0.00\ [-4.38,\,4.38]$	$+5.72\ [-1.68,\,13.47]$
All-G $-$ All-I	$-0.30\ [-0.71,\,0.10]$	$-1.01\ [-2.69,\,0.34]$	$-1.68\ [-5.39,\,1.68]$
Count-G $-$ Count-I	$-0.20\ [-0.61,\,0.17]$	$-1.68\ [-3.70,\,-0.34]$	$+0.67\ [-3.03,\,4.04]$
Consistent-I $-$ Count-I	$-0.27\ [-0.61,\,0.03]$	$-1.35\ [-2.69,\,-0.34]$	$+1.01\ [-3.03,\,5.05]$
Consistent-G $-$ Count-G	$+0.07\ [-0.40,\,0.54]$	$+0.34\ [-0.67,\,1.68]$	$-0.34\ [-3.03,\,2.36]$
Consistent-G $-$ Consistent-I	$+0.13\ [-0.24,\,0.51]$	$+0.00\ [-1.01,\,1.01]$	$-0.67\ [-4.04,\,2.69]$

Appendix

Table 9: Descriptive 95% paired-complex percentile bootstrap intervals for the predeclared all-complex contrasts. Entries are percentage-point differences, with the first policy minus the second; intervals use 10,000 resamples and are not seed-level significance intervals.

Figure 4: Integrated summary of the post-training endpoint separation. (a) The starting model and broad intrinsic continuation differ in fixed score loss, aggregate best-of-10 recovery, and confidence-selected Top-1 success; the starting model is one checkpoint and the broad values are three-seed means. (b) Seed-matched changes for the factorial policies are shown relative to broad intrinsic continuation; each dot is one training seed, and the broad/focused mixture is shown separately because it is not a factorial cell. (c) Source-to-broad changes are shown by torsional flexibility stratum. The figure re-expresses the quantities reported in the matched-study and flexibility summaries.

Figure 5: Matched post-training outcomes. (a) Fixed score loss; (b) aggregate generated-pose, best-of-10, and confidence-selected Top-1 success; and (c) the same docking endpoints for the $1$ – $5$ and $6+$ torsion strata. Bars show the starting model and three-seed policy means; error bars show seed standard deviations where available. The common PDBBind evaluation uses ten poses per complex and a $5\,\text{\AA}$ threshold.

Policy	Stratum	$n$	Generated	Best-of-10	Top-1
Starting model	all	99	45.76	86.87	34.34
Broad intrinsic	all	99	47.14	86.87	40.07
Mixture	all	99	46.94	86.20	39.06
Starting model	$1$ – $5$ torsions	64	53.91	90.63	42.19
Broad intrinsic	$1$ – $5$ torsions	64	56.35	92.19	47.92
Mixture	$1$ – $5$ torsions	64	56.20	92.19	50.00

Appendix

Table 10: Flexibility-stratified percentages on the 99-complex PDBBind test evaluation at the

5\,\text{\AA}

threshold. Values use ten generated poses per complex; the four zero-torsion complexes are retained in the all-complex result but are not interpreted as a subgroup.

Policy	Generated $<2\,\text{\AA}$	Best-of-10 $<2\,\text{\AA}$	Top-1 $<2\,\text{\AA}$
Starting model	17.98	49.49	14.14
Broad intrinsic	18.99	48.82	14.81
Broad Cartesian	19.12	49.16	15.15
Count-matched intrinsic	18.89	49.16	14.14
Count-matched Cartesian	19.02	49.49	15.49
Consistent intrinsic	18.92	49.49	15.15

Appendix

Table 11: Supporting

2\,\text{\AA}

docking endpoints on the 99-complex PDBBind test evaluation. Values are percentages; trained-policy values are three-seed means. The

5\,\text{\AA}

threshold remains primary.

Policy	Threshold	Random reference	Confidence Top-1	Top-1 minus reference
Starting model	$2\,\text{\AA}$	17.98	14.14	$-3.84$
Broad intrinsic	$2\,\text{\AA}$	18.99	14.81	$-4.18$
Starting model	$5\,\text{\AA}$	45.76	34.34	$-11.42$
Broad intrinsic	$5\,\text{\AA}$	47.14	40.07	$-7.07$

Appendix

Table 12: Confidence-selected success compared with the expected success from uniformly selecting one of ten generated poses. Values are percentages; the random reference equals generated-pose success.

Molecular Docking

See all 16 papers

Curriculum Learning for Generative Flexible Docking
Oct 20, 2026Wenyin Zhou, Hossein AzizpourMolecular Docking Structure-Based Drug Design
KTH Royal Institute of Technology · Science for Life Laboratory
Complete Donor-Set Guidance for Frozen Diffusion Docking Models
Nov 1, 2026Hyunwoo Kim, Insoo Park, Ziseok Lee +3Molecular Docking Diffusion Model Guidance
Seoul National University
FuseDiff: Symmetry-Preserving Joint Diffusion for Dual-Target Structure-Based Drug Design
Mar 5, 2026Jianliang Wu, Anjie Qiao, Zhen Wang +2Structure-Based Drug Design Drug Discovery
School of Computer Science and Engineering, Sun Yat-sen University, Guangzhou, China · School of Computer Science and Engineering, Guangdong Province Key Laboratory of Computational Science, Sun Yat-sen University, Guangzhou, China · Gaoling School of Artificial Intelligence, Renmin University of China, Beijing, China +1

Oct 20, 2026J/K move · Enter open · S save

Curriculum Learning for Generative Flexible Docking

Wenyin Zhou, Hossein Azizpour

KTH Royal Institute of Technology · Science for Life Laboratory

Molecular Docking Structure-Based Drug Design Curriculum Learning

Open paper AskPDF arXiv

Flexible docking is a highly practical task, aiming to predict the joint configuration of a protein-ligand complex from a protein apo structure and a ligand SMILES string. Diffusion and flow-based generative models have made this task tractable by modeling the posterior distribution over the ligand pose and the protein conformation conditioned on the apo input. The central challenge, however, remains the long-tailed apo-to-holo structure shift, where a small fraction of protein undergo large conformational changes. In this work, motivated by the practical observation that downweighting large shift examples improves generalization, and by the mixed nature of the long tail consisting of genuinely difficult examples and noisy ones, we propose to use curriculum learning for this task, gradually incorporating large conformational change examples during training. With a linear pacing function, we show that an existing method can be strongly enhanced. On the PDBBind ESMFold apo benchmark, with relaxation and confidence-based pose selection, the resulting model GCLDock achieves a 46.52% top-1 success rate (ligand RMSD less than 2A) and 40.26% when additionally requiring PoseBusters validity, outperforming existing baselines.

Abstract

Molecular Docking

Explore similar work

Curriculum Learning for Generative Flexible Docking

Complete Donor-Set Guidance for Frozen Diffusion Docking Models

FuseDiff: Symmetry-Preserving Joint Diffusion for Dual-Target Structure-Based Drug Design