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

The 30-second take

• What: A fixed-budget DiffDock study compares broad continuation with torsional supervision aimed at Cartesian consequence or conditional target consistency, plus a mixture, and finds broad continuation had the highest mean Top-1 success at under 5 angstroms among trained policies.
• Why it matters: Docking diffusion models are tempted to spend scarce post-training on the errors that look structurally serious. If that focus does not improve returned poses, the default should stay broad coverage of the score field.
• Who should care: Molecular-docking ML groups, anyone fine-tuning DiffDock-class models, and reviewers who treat geometric-relevance weighting as obviously better.

What the paper actually did

Diffusion-based molecular docking has to learn a corrective score field across a broad corruption distribution, even though individual pose errors differ in structural consequence. That creates a post-training conflict between keeping broad score-field coverage and concentrating on corrections that look 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 splits fixed score fidelity, generated-pose quality, best-of-10 sampling capacity, and confidence-selected Top-1 docking performance.

At under 5 angstroms, broad continuation had the highest mean Top-1 success among trained policies. Focused or mixed supervision showed no consistent advantage under the tested budget. The starting model had better fixed score fidelity, while the evaluated broad-continuation model had higher mean Top-1 success with the same aggregate best-of-10 recovery. Focusing on structurally consequential or more consistently specified corrections did not consistently improve returned-pose success over broad continuation.

What makes this disruptive

If the core claim holds, a fashionable training prior is wrong for this budget: weighting the "important" docking errors is not a free improvement. The scarce resource is post-training compute, and spending it on geometric relevance failed to beat more of the original objective.

The distinctive result is a negative one with a clean split: starting checkpoint wins on fixed score fidelity; broad continuation wins mean Top-1; focused and mixed runs do not consistently help; best-of-10 recovery stays comparable. That is more useful than another docking leaderboard bump.

This week's editorial ranking placed the preprint ninth of twenty, with a score of 78. That is triage for the Disruptive Concepts brief, not a peer-review verdict. Treat it as a post-training ablation, not as a new docking architecture.

Why it matters (outside the lab)

Abundance lens (today's luxuries toward tomorrow's defaults): reliable structure-based design still sits in well-funded loops. If post-training recipes that look geometrically clever do not move Top-1 poses, compute should stay on broad score coverage until someone shows otherwise.

Near-term, the preprint is a caution for DiffDock fine-tunes. Medium-term, other budgets, other docking diffusion models, and whether confidence selection hides sampling gains decide how far the negative result travels.

No calendar date is implied. Disruptive Concepts, which publishes these explainers free for humans, reads this as a possible brick toward making docking training less cargo-cult, not as a promise of better drugs.

Limitations & open questions

This is a preprint ablation on DiffDock, not a new method that wins the field. The abstract reports a fixed budget, an under-5-angstrom Top-1 comparison, and no consistent focused advantage. We have not re-run the policies.

The abstract does not report dataset names, the exact budget, variance, or whether a larger focused budget would flip the result. Best-of-10 matching aggregate recovery means the sampling capacity may be unchanged while Top-1 selection moves. "Geometric relevance" is operationalized as Cartesian consequence or conditional target consistency; other focus schemes are untested.

Abundance is not automatic. A negative ablation does not demonetize docking on a fixed date. Evaluation choices still sit between this study and a default recipe.