AF3Complex: PPI Modeling with pIS Scoring

• AF3Complex is a protein–protein interaction predictor that uses sequence-only input to model viral–host complexes and calculates a predicted Interface-Similarity (pIS) score to assess native-like binding.
• It replaces the traditional ipTM metric with a pIS metric, achieving a recall of about 69% on 13 curated virus–host PPIs while revealing sensitivity to receptor context and missing biological nuances.
• The evaluation highlights limitations such as dataset bias, absence of membrane and glycan context, and weak calibration in mutational stress tests, suggesting its use as a preliminary flag for further wet-lab validation.

Searching arXiv for “AF3Complex” and closely related entries to ground the article in current papers. AF3Complex is a protein–protein interaction (PPI)-focused derivative of AlphaFold 3 that attempts to model protein–protein complexes and score their interfaces more directly. In the biosecurity evaluation reported in "Resilient Biosecurity in the Era of AI-Enabled Bioweapons" (Feldman et al., 30 Aug 2025), it builds on AlphaFold 3’s multimer modeling but replaces AlphaFold 3’s interface-predicted TM confidence with a predicted Interface-Similarity score (pIS), a confidence quantity conceptually tied to the IS-score family of interface similarity metrics from docking and complex benchmarking. In the same supplied arXiv literature, however, the label "AF3Complex" is also reused for unrelated constructs in numerical analysis, finite element complexes, and aircraft conflict resolution, so the term is best treated as context-dependent rather than uniquely standardized.

1. Core definition in structural biology

Within the virus–host PPI setting, AF3Complex is used as a sequence-to-complex predictor. The model is run from sequence alone for both the viral protein and the human receptor, with no template structures or pre-supplied docking poses, and it returns a pIS score for each modeled complex. In the cited evaluation, binding is declared when $pIS \ge 0.38$; for comparison, AlphaFold 3 is evaluated with $ipTM \ge 0.6$, and SpatialPPIv2 with probability $\ge 0.5$ (Feldman et al., 30 Aug 2025).

The methodological distinction is not merely cosmetic. By substituting pIS for ipTM, AF3Complex is intended to estimate how “native-like” a modeled interface is, rather than relying on AlphaFold 3’s default interface confidence proxy. This places the method closer to interface-quality assessment as used in docking and complex benchmarking. A plausible implication is that AF3Complex is optimized for the plausibility of the interface itself, rather than for multimer confidence in a broader sense, although the reported evaluation shows that this shift does not by itself yield reliable screening behavior on biologically important viral interactions.

2. Evaluation design and benchmark construction

The benchmark consists of thirteen well-characterized virus–host PPIs selected from cases where complexes have been resolved or curated in the Protein Data Bank. The ground-truth references include SARS-CoV-2 Spike–ACE2 (6M0J), MERS (4L72), HIV (1GC1), adenovirus (1F5W), HSV–HVEM (1JMA), HSV–Nectin-1 (3SKU), hepatitis B (8HRX), Epstein–Barr (8SM0), measles virus (3INB and 4GJT), Ebola (5F1B), Nipah (2VSM), and Chikungunya (6JO8). AF3Complex and AlphaFold 3 receive only the two sequences, whereas SpatialPPIv2 uses structures precomputed by AlphaFold 3 (Feldman et al., 30 Aug 2025).

The study also includes a mutational stress test on four experimentally validated SARS-CoV-2 Spike RBD mutants from Moulana et al. (Nature Communications, 2022). Two variants exhibit more than tenfold increases in ACE2-binding affinity, reflected as negative $\Delta \log_{10} KD$ values, and two exhibit substantial decreases while retaining functional binding. The authors additionally note that many of the tested virus–host interactions appear in the training sets of the evaluated models, including AlphaFold 3 and AF3Complex, yet several known interactions are still missed. This is significant because it frames the observed failures not only as out-of-distribution breakdowns but also as a generalization and calibration problem on familiar targets.

3. Recognition performance on curated virus–host PPIs

On the thirteen positive virus–host PPIs, AF3Complex identifies 9 of 13 interactions and fails on 4. Its failures are Chikungunya (6JO8), hepatitis B (8HRX), HSV–HVEM (1JMA), and the non-Edmonston measles strain (4GJT). It succeeds on adenovirus (1F5W), Ebola (5F1B), Epstein–Barr (8SM0), HIV (1GC1), HSV–Nectin-1 (3SKU), measles (Edmonston strain, 3INB), MERS (4L72), Nipah (2VSM), and SARS-CoV-2 Spike–ACE2 (6M0J) (Feldman et al., 30 Aug 2025).

Because the test set contains only positives, the study reports recall but not precision, specificity, AUROC, or F1. For AF3Complex, recall is

$$\mathrm{Recall} = \frac{TP}{TP+FN} = \frac{9}{13} \approx 0.69,$$

with misidentifications of about $30\%$.

Model	Decision threshold	Benchmark outcome
AF3Complex	$pIS \ge 0.38$	9/13 correct, 4/13 missed
AlphaFold 3	$ipTM \ge 0.6$	7/13 correct, 6/13 missed
SpatialPPIv2	probability $\ge 0.5$	8/13 correct, 5/13 missed

Among the individual cases, SARS-CoV-2 Spike–ACE2 is notable because AF3Complex correctly classifies the wildtype interaction as binding, while AlphaFold 3 and SpatialPPIv2 do not meet their own thresholds. Hepatitis B is notable for the opposite reason: all three models fail. The HSV receptor-specificity pattern is also mixed, with AF3Complex succeeding on Nectin-1 but not on HVEM, which underscores receptor-specific sensitivity and probable domain or context effects in interface modeling.

4. Mutational stress tests and confidence calibration

AF3Complex fails to flag any of the four experimentally validated SARS-CoV-2 Spike RBD mutants as binders, including the two variants with stronger ACE2 binding. The reported pIS scores are 0.11, 0.06, 0.06, and 0.05, all below the binding threshold of 0.38. Relative to the wildtype Spike–ACE2 baseline, the model’s confidence decreases by roughly 0.47–0.53 points, with an inferred wildtype baseline pIS of about 0.58 (Feldman et al., 30 Aug 2025).

This behavior is important because it does not track the direction of the true affinity change. Two mutants with $\Delta \log_{10} KD \approx -1.04$ and $ipTM \ge 0.6$0 are experimentally stronger binders, yet AF3Complex lowers confidence on them; the two weaker but still binding mutants with $ipTM \ge 0.6$1 and $ipTM \ge 0.6$2 are also scored as non-binding. The companion comparison sharpens the calibration problem: SpatialPPIv2 and AlphaFold 3 slightly increase confidence on both beneficial and detrimental mutations, whereas AF3Complex consistently decreases confidence, and none of the three models tracks the measured affinity changes. The result directly supports the study’s claim of weak calibration under sequence perturbation and limited sensitivity to combinatorial epistasis.

5. Failure modes, scope limits, and biosecurity relevance

The reported failure analysis identifies several factors that are observed or likely to impair AF3Complex. These include viral membrane proteins and heavily glycosylated interfaces, such as Spike, where missing glycans, cofactors, or membrane context can invalidate docking poses or confidence; disordered regions and conformational flexibility, which are not captured from sequence alone; stoichiometry and assembly dependence, because many viral complexes rely on oligomeric states or accessory proteins; and mutational epistasis, where multi-mutation effects on binding are non-additive and poorly reflected in single-structure confidence metrics (Feldman et al., 30 Aug 2025).

The paper also emphasizes dataset bias and scarcity. Public corpora contain relatively few high-quality, annotated viral–host complexes compared with the vast number of non-viral proteins, which yields poor learning of viral interface idiosyncrasies. More broadly, purely structural predictors omit receptor glycosylation, membrane topology, proteolytic activation, and competitive binding in vivo. The study therefore concludes that AF3Complex, although best among the three tools evaluated, is not reliable enough to serve as an inference-time safety filter: its approximately $ipTM \ge 0.6$3 miss rate on the curated benchmark and complete failure on all four experimentally validated SARS-CoV-2 mutants create a high false-negative risk, which is precisely the most dangerous error mode for harmful-output screening. The proposed strategic response is a shift from prediction-centric defenses to response-oriented infrastructure, including rapid experimental validation, adaptable biomanufacturing, and regulatory pathways operating at AI time scales.

The same discussion also outlines possible improvement directions for AF3Complex-based screening: expanding and rebalancing viral–host PPI datasets, incorporating membrane context and glycan-aware modeling, calibrating thresholds against richer negative sets and more diverse viral domains, using ensembles of orthogonal predictors, and treating any predicted binder as a flag for wet-lab triage rather than definitive evidence. In that framing, AF3Complex becomes one weak signal among several rather than a gatekeeper.

6. Nomenclature ambiguity and other arXiv uses of “AF3Complex”

A recurring misconception would be to treat AF3Complex as a unique, domain-stable proper name. In the supplied arXiv literature, the label is explicitly reused for unrelated technical objects. One paper maps the term to a fast multipole method for complex-coordinate geometries, where a hierarchical tree is built on the real parts of complex point locations and linear-time complexity is retained for standard Laplace and Helmholtz layer potentials (Goodwill et al., 5 Sep 2025). Another maps it to three-dimensional conforming finite element divdiv complexes on cuboid and tetrahedral meshes, emphasizing exactness, enhanced smoothness, commuting projections, and implementable degrees of freedom (Hu et al., 2022). A third uses the label for a complex-number formulation of aircraft conflict resolution, with one binary variable per aircraft pair, disjunctive linear separation constraints, and tight convex relaxations aimed at global optimality (Rey et al., 2017).

This terminological overloading matters for bibliographic retrieval and for technical discussion. In structural biology, AF3Complex refers to a PPI-focused derivative of AlphaFold 3 with pIS-based interface scoring. In numerical analysis, finite element theory, and operations research, the same string denotes entirely different objects. For that reason, any scholarly use of the term benefits from immediate domain qualification, such as “AF3Complex for viral–host PPI prediction” or “AF3Complex in the complex-coordinate FMM sense,” to prevent cross-domain ambiguity.