AlphaFold 3 Confidence Scores Enter Drug Prioritization Workflows
AlphaFold 3 is moving beyond structure prediction and into the more consequential problem of drug prioritization. In this preprint, investigators benchmarked its interaction confidence scores against BANDOCK+, an experimentally informed method, using approved drugs and human protein libraries. The central question was not whether AF3 can model protein-ligand complexes in principle, but whether its internal confidence metric carries enough signal to help rank candidate therapies at proteome scale.
The answer was cautiously affirmative. AF3 ipTM achieved similar aggregate disease ranking performance to BANDOCK+, although the performance varied across individual indications. For computational drug discovery, that is a meaningful result. A model designed to predict molecular interaction confidence can still behave like a ranking tool when it is pushed across a large human proteome, yet the signal is not uniform enough to treat every disease area as equivalent.
Why Interaction Confidence Matters in a Proteome-Scale Screen
The appeal of confidence-based prioritization is straightforward. Drug repurposing and target nomination require some way to narrow a vast protein-drug search space, and structural scoring remains one of the few approaches that can be deployed across thousands of proteins without an immediate dependency on fresh wet-lab data. AF3 adds a new layer to that process by estimating the plausibility of protein-ligand complexes, while BANDOCK+ represents a comparator grounded in experimental information.
The practical implication is not that one method replaces the other. It is that AF3 confidence may serve as an additional computational filter, especially when access to binding data is sparse or incomplete. Proteome-scale ranking is unforgiving: even a modest enrichment of plausible targets can alter which compounds move forward for downstream validation. In that setting, a score that aligns with disease ranking performance seen in a benchmarked, experimentally informed framework deserves attention.
The biology behind the strategy is intuitive. Proteins do not function as isolated atoms of disease; they occupy networks, pathways, and structural states that determine whether a ligand is likely to engage a clinically relevant target. A confidence score that captures the stability or plausibility of a predicted complex may therefore preserve enough functional signal to sort candidate drug-target pairs in a way that is useful for indication-level triage.
Agreement at the Aggregate Level, Divergence at the Indication Level
The strongest finding in this study is also the most restrained. AF3 ipTM matched BANDOCK+ at the aggregate disease-ranking level, but individual indications did not all behave the same way. That pattern matters because it argues against overgeneralizing from a single global benchmark. Some disease contexts may be more structurally tractable, more chemically compatible, or simply better represented by the proteins and approved drugs included in the analysis.
A one-size-fits-all interpretation would overstate the result. AF3 confidence can be informative, but the usefulness of that information appears to depend on the specific indication being ranked. For developers and computational biologists, this means the metric is promising as a broad prioritization signal, while still requiring indication-specific calibration before it can be treated as a stand-alone decision tool.
This heterogeneity is not surprising. Drug prioritization is shaped by target biology, protein family behavior, ligand class, and the degree to which the benchmark set reflects real pharmacology. A method that performs well across the full dataset can still vary in usefulness when it is applied to a narrow therapeutic area with unusual binding constraints or limited structural precedent.
What the Benchmark Suggests About AF3’s Clinical Utility
The preprint positions AF3 interaction confidence as a source of information for proteome-scale drug prioritization, not as a final arbiter of therapeutic success. That distinction is important at the bedside and in translational research alike. A high-confidence interaction does not prove efficacy, and a low-confidence interaction does not rule out a drug’s clinical value. Binding is only one step in a long chain that includes permeability, metabolism, pathway dependence, toxicity, and actual disease modification.
Still, the benchmark has real translational weight. Approved drugs were used alongside human protein libraries, which means the system was tested in a setting that approximates repurposing rather than purely theoretical docking. That choice helps connect the analysis to familiar therapeutic chemistry and to human protein targets that matter in disease. For teams building discovery pipelines, a method that reaches comparable aggregate ranking performance to BANDOCK+ may justify inclusion as one more layer in a consensus framework.
There is also a workflow implication. Proteome-scale screening can be overwhelmed by the sheer number of candidate interactions. Confidence scores that reduce false leads, even imperfectly, can save experimental time and focus follow-up assays on the most plausible candidates. In that sense, the value of AF3 may lie less in absolute prediction and more in its ability to sharpen the front end of a prioritization funnel.
Preprint Status Leaves Key Validation Questions Open
This work remains a preprint, and that status shapes how the findings should be used. Peer review has not yet certified the analysis, and the abstract provides limited detail on the exact benchmarking design, disease set composition, or the full range of ranking outcomes. Those missing details matter for interpreting generalizability. Without them, the safest reading is that AF3 ipTM is a credible candidate signal for large-scale prioritization, not a finished clinical or translational standard.
Several questions now follow from the benchmark itself. How stable is AF3 performance across protein classes? Which indication types benefit most from confidence-based ranking, and which are poorly served? Can a hybrid approach combining AF3 confidence with experimentally informed methods outperform either alone in a reproducible way? Those are the kinds of questions that will determine whether interaction confidence becomes a routine part of repurposing pipelines or remains an attractive but limited adjunct.
For now, the study establishes a narrower but useful point: AF3 interaction confidence can track disease ranking performance at a scale large enough to matter, and it does so with enough consistency to merit further validation against experimentally anchored methods like BANDOCK+.