Geometric Screening: Concepts & Applications

• Geometric screening is the study of how explicit geometric structures, such as curves, safe sets, and projection operations, characterize screening phenomena across various domains.
• In convex optimization, geometric screening creates certified convex regions that safely discard inactive predictors, ensuring optimal solutions without false rejections.
• Applications in molecular screening and spatial prediction utilize geometric deep learning and observation configurations to enhance ranking accuracy and computational efficiency.

“Geometric screening” is an Editor’s term for a set of research programs in which a screening phenomenon is characterized through explicit geometric structure rather than only through scalar summaries or black-box computation. In the supplied literature, the phrase spans several technically distinct domains: diagnostic testing, where sensitivity and specificity generate curves on a unit square and yield geometric definitions of the positive likelihood ratio; convex optimization, where safe regions in dual or primal space certify that variables, groups, or samples can be discarded without changing the optimum; molecular virtual screening, where geometric deep learning jointly predicts protein–ligand structure and screening scores; and condensed-matter or continuum theories, where induced quadrupoles, dipoles, or host-lattice deformations screen fields or dipole moments. The common motif is that screening is represented by shape, projection, curvature, invariant points, safe sets, or emergent length scales rather than by a purely symbolic criterion (Balayla, 2020).

1. Diagnostic-test geometry and the screening plane

In the medical-screening formulation, the basic object is the unit square $S=[0,1]\times[0,1]$, with horizontal axis equal to the pre-test probability or disease prevalence $\phi\in[0,1]$ and vertical axis equal to the positive predictive value $\rho(\phi)$. A screening test with sensitivity $a$ and specificity $b$ induces the continuous curve

$\rho(\phi)=\frac{a\phi}{a\phi+(1-b)(1-\phi)}.$

This curve always passes through the origin invariant $(0,0)$ and the endpoint invariant $(1,1)$, two fixed points shared by every screening curve (Balayla, 2020).

The same work defines the prevalence threshold $\phi_e$ as the abscissa of the inflection point, described as the point of greatest curvature of the screening curve, with

$$\phi_e=\frac{\sqrt{a(1-b)}+b-1}{\varepsilon-1},\qquad \varepsilon=a+b.$$

Below $\phi\in[0,1]$0, false positives dominate the denominator of $\phi\in[0,1]$1; above $\phi\in[0,1]$2, the true-positive term $\phi\in[0,1]$3 dominates. The paper further states the fundamental theorem of screening: $\phi\in[0,1]$4 Here $\phi\in[0,1]$5 is the screening coefficient. The same source emphasizes that identical values of $\phi\in[0,1]$6 do not determine a unique curve shape, because different $\phi\in[0,1]$7 pairs with the same sum can generate different geometries and different areas under $\phi\in[0,1]$8 (Balayla, 2020).

A first geometric definition of the positive likelihood ratio follows from the line joining the origin to the threshold point $\phi\in[0,1]$9. If $\rho(\phi)$0 is the small angle between this line and the vertical axis, then

$\rho(\phi)$1

A second definition uses an arbitrary point $\rho(\phi)$2 on the screening curve and the lines from $\rho(\phi)$3 to the two invariant points. If $\rho(\phi)$4 joins $\rho(\phi)$5 to $\rho(\phi)$6 and $\rho(\phi)$7 joins $\rho(\phi)$8 to $\rho(\phi)$9, then

$a$0

These constructions are proposed as a formal way to distinguish tests with the same $a$1 but different geometries across the prevalence spectrum (Balayla, 2020).

2. Safe geometric screening in convex optimization

A second major meaning of geometric screening appears in sparse optimization and empirical risk minimization, where geometry is used to eliminate variables, groups, or samples before or during optimization. In the Lasso, the primal problem

$a$2

has a dual feasible set

$a$3

described as a centrally symmetric polytope. The dual optimum $a$4 is the Euclidean projection of $a$5 onto that polytope. Screening then proceeds by constructing a convex safe region $a$6 known to contain $a$7; if the maximal possible correlation of a feature with any point in $a$8 stays strictly below the dual constraint boundary, the corresponding coefficient is guaranteed to be zero and can be discarded (Xiang et al., 2014).

The survey on Lasso screening organizes this logic through sphere, dome, and two-hyperplane tests. For a Euclidean ball $a$9, the support function is

$b$0

yielding the safe-discard condition

$b$1

The same survey presents sequential screening, in which a solution at $b$2 provides a tighter safe region for $b$3, and describes Data-Adaptive Sequential Screening as a way to control the number of steps (Xiang et al., 2014).

The dual-polytope-projection line of work makes the same geometry more explicit. Because the projection onto a nonempty closed convex set is unique and nonexpansive, one obtains

$b$4

This places $b$5 inside a ball around the known $b$6, which leads to the DPP rule

$b$7

The same paper extends the construction to group Lasso by replacing scalar feature bounds with operator-norm bounds on groups (Wang et al., 2012).

A related but broader framework treats screening rules for convex problems via the duality gap. For the minimum enclosing ball, if $b$8 is a primal center, $b$9 a primal radius, and $\rho(\phi)=\frac{a\phi}{a\phi+(1-b)(1-\phi)}.$0 a dual-feasible simplex point, the duality gap satisfies

$\rho(\phi)=\frac{a\phi}{a\phi+(1-b)(1-\phi)}.$1

The corresponding safe rule is

$\rho(\phi)=\frac{a\phi}{a\phi+(1-b)(1-\phi)}.$2

which certifies that point $\rho(\phi)=\frac{a\phi}{a\phi+(1-b)(1-\phi)}.$3 lies strictly in the interior of the optimal ball and cannot be a support point (Raj et al., 2016).

Geometric screening in ERM uses a different safe set: an ellipsoid $\rho(\phi)=\frac{a\phi}{a\phi+(1-b)(1-\phi)}.$4 guaranteed to contain the primal minimizer $\rho(\phi)=\frac{a\phi}{a\phi+(1-b)(1-\phi)}.$5. If a loss has a flat region $\rho(\phi)=\frac{a\phi}{a\phi+(1-b)(1-\phi)}.$6, and if for sample $\rho(\phi)=\frac{a\phi}{a\phi+(1-b)(1-\phi)}.$7

$\rho(\phi)=\frac{a\phi}{a\phi+(1-b)(1-\phi)}.$8

then the optimal dual coordinate satisfies $\rho(\phi)=\frac{a\phi}{a\phi+(1-b)(1-\phi)}.$9, so the sample can be discarded safely. This geometry is coupled to “safe loss functions” and dual sparsity-inducing regularization (Mialon et al., 2019).

These optimization papers share a strict notion of safety: discarded objects are certified to be inactive at the exact optimum. A common misconception is to conflate such rules with heuristic pruning. The supplied material distinguishes safe rules, which never discard active predictors or samples under their assumptions, from heuristic rules such as the Lasso “strong rules,” for which false rejections can occur with very low probability (Xiang et al., 2014).

3. Geometric screening in molecular virtual screening

In structure-based drug discovery, geometric screening refers to models that encode molecular geometry directly into the screening pipeline rather than relying only on docking search plus ex post scoring. LigPose represents each protein–ligand pair as a graph whose nodes are pocket and ligand atoms and whose edges form a complete, undirected graph. Edge features include Euclidean distance encoded via a small RBF basis together with covalent-bond indicators, while coordinate updates are SE(3)-equivariant. The model is trained on three supervised tasks—pose optimization, binding affinity, and binding probability—and, for large-scale screening, outputs predicted ligand pose $(0,0)$0, predicted affinity $(0,0)$1, and predicted binding probability $(0,0)$2. The screening score is defined as

$(0,0)$3

Ligands are ranked by $(0,0)$4 (He et al., 2024).

The same source reports virtual-screening metrics including EF@1%, EF@5%, success rates at top-1%, top-5%, and top-10% of the ranked library on CASF-2016, and AUC on larger ligand sets. On CASF-2016, LigPose achieves $(0,0)$5 versus $(0,0)$6 for RTMScore and Success@1% $(0,0)$7 versus $(0,0)$8. On a SARS-CoV-2 $(0,0)$9 repurposing set, the AUC is $(1,1)$0 versus $(1,1)$1 for ImageMol. The paper also reports wall-clock inference of approximately $(1,1)$2–$(1,1)$3 s per complex on GPU and speedups of $(1,1)$4–$(1,1)$5 on CPU and $(1,1)$6–$(1,1)$7 on GPU versus Smina, Glide, and Vina (He et al., 2024).

Dockformer uses a different multimodal geometric representation. It combines one-dimensional atom encodings, two-dimensional graph pair features based on shortest-path distance and path-averaged edge features, and three-dimensional position and distance embeddings. Pair-biased transformer blocks fuse ligand and pocket features, and a structure module updates ligand coordinates according to

$(1,1)$8

Dockformer also predicts a confidence measure and is evaluated on PDBbind core and PoseBusters. The reported success rates are $(1,1)$9 and $\phi_e$0, respectively, with more than a $\phi_e$1-fold increase in inference speed relative to traditional methods (Yang et al., 2024).

A plausible implication is that, in this literature, “geometric screening” names not merely a ranking task but an end-to-end geometric inference problem: the same architecture learns structure, interaction, and screening. The supplied sources explicitly contrast this with classical docking, where geometry enters primarily through repeated evaluation of per-pose energy functions and global search heuristics (He et al., 2024).

4. Screening and antiscreening in materials and continuum mechanics

In materials physics, geometric screening denotes the modification of fields by induced geometric objects or by geometry-dependent host response. One line of work concerns endohedral nanotubes containing polar molecules. If an isolated molecule has bare dipole moment $\phi_e$2, the effective dipole of the composite is

$\phi_e$3

Here $\phi_e$4 denotes screening and $\phi_e$5 denotes antiscreening. The effective dipole is computed from the total charge distribution,

$\phi_e$6

and the induced electronic contribution is obtained from the differential charge density $\phi_e$7 (Silvestrelli et al., 2024).

The same study shows that geometry is decisive. Covalent carbon nanotubes exhibit strong screening, partially ionic nanotubes show reduced screening, and some ionic LiF nanotubes exhibit antiscreening. The cross-section is crucial: octagonal LiF nanotubes can show antiscreening, whereas dodecagonal LiF nanotubes are always characterized by a reduction of the total dipole moment. For example, $\phi_e$8 oct-6 has $\phi_e$9 D and $$\phi_e=\frac{\sqrt{a(1-b)}+b-1}{\varepsilon-1},\qquad \varepsilon=a+b.$$0, while $$\phi_e=\frac{\sqrt{a(1-b)}+b-1}{\varepsilon-1},\qquad \varepsilon=a+b.$$1 dod-6 has $$\phi_e=\frac{\sqrt{a(1-b)}+b-1}{\varepsilon-1},\qquad \varepsilon=a+b.$$2 D and $$\phi_e=\frac{\sqrt{a(1-b)}+b-1}{\varepsilon-1},\qquad \varepsilon=a+b.$$3. The paper attributes these differences to the interplay of electronic polarization, radial expansion or contraction, axial displacements, and tube geometry (Silvestrelli et al., 2024).

A distinct continuum-mechanics literature models screening through plasticity-generated geometric charges. In two-dimensional solids, local stress-relaxation mechanisms are treated as quadrupolar in nature, and the theory introduces a hierarchy of screening modes: quadrupole screening, dipole screening, and monopole screening. The unified screened Airy-stress equation is

$$\phi_e=\frac{\sqrt{a(1-b)}+b-1}{\varepsilon-1},\qquad \varepsilon=a+b.$$4

where $$\phi_e=\frac{\sqrt{a(1-b)}+b-1}{\varepsilon-1},\qquad \varepsilon=a+b.$$5 and $$\phi_e=\frac{\sqrt{a(1-b)}+b-1}{\varepsilon-1},\qquad \varepsilon=a+b.$$6 are internal length scales associated with dipole and monopole screening, respectively. The paper argues that the dipole-screened regime is mechanically identical to the hexatic phase, suggesting that “hexatic” can be defined mechanically rather than structurally (Livne et al., 2023).

Dynamic screening by plasticity in amorphous solids extends this idea to oscillatory loading. A screening stress $$\phi_e=\frac{\sqrt{a(1-b)}+b-1}{\varepsilon-1},\qquad \varepsilon=a+b.$$7 generated by plastic quadrupoles produces, after elimination of $$\phi_e=\frac{\sqrt{a(1-b)}+b-1}{\varepsilon-1},\qquad \varepsilon=a+b.$$8, a mass-like term $$\phi_e=\frac{\sqrt{a(1-b)}+b-1}{\varepsilon-1},\qquad \varepsilon=a+b.$$9. The resulting equation is

$\phi\in[0,1]$00

with screening length $\phi\in[0,1]$01. In circular geometry the radial displacement satisfies a Bessel equation, and when $\phi\in[0,1]$02 the displacement decays exponentially rather than algebraically (Hentschel et al., 2023).

Experimental and numerical work on granular matter reports a related strong-screening equation,

$\phi\in[0,1]$03

with screening length

$\phi\in[0,1]$04

In the strong-screening regime the displacement field becomes negative in an intermediate annulus before returning to zero at the boundary, and the extracted $\phi\in[0,1]$05 from experiment agrees within errors with DEM simulation (Mondal et al., 2021).

A recurrent misconception in this mechanics literature is to treat screening as merely a renormalization of elastic constants. The supplied sources separate weak screening, which renormalizes moduli, from strong or anomalous screening, which introduces a finite screening length and qualitatively changes Green’s functions and displacement profiles (Mondal et al., 2021).

5. Screening in spatial prediction

In spatial statistics, the “screening effect” is the empirical and theoretical phenomenon that nearby observations dominate optimal linear prediction, so that distant observations contribute negligibly once sufficiently close observations are present. For a mean-zero stationary Gaussian process $\phi\in[0,1]$06 with spectral density $\phi\in[0,1]$07, if $\phi\in[0,1]$08 denotes a near set and $\phi\in[0,1]$09 a far set, the asymptotic screening effect is defined by

$\phi\in[0,1]$10

This makes the concept geometric through the configuration of observation locations and spectral through the behavior of $\phi\in[0,1]$11 at high frequencies (Stein, 2012).

The key spectral condition in the supplied source is

$\phi\in[0,1]$12

The paper conjectures that, together with a geometric derivative-prediction condition on the near set, this implies asymptotic screening for broad observation schemes. Two cases are proved. In one dimension, if

$\phi\in[0,1]$13

and the high-frequency condition holds, then screening holds for any finite near set and any far set. In two dimensions, if the process is not mean-square differentiable in any direction and the near and far sets have the specified configuration $\phi\in[0,1]$14 and $\phi\in[0,1]$15, then screening also holds (Stein, 2012).

The same paper gives explicit failure modes. Screening may fail when the spectral density oscillates at high frequency, as in the triangular covariance example, or when the near-point geometry is too sparse to estimate relevant derivatives, as in the once-differentiable Matérn example with only one nearby observation. The practical recommendation is accordingly twofold: use models whose spectral density changes slowly at high frequencies, and match the geometry of observation locations to the process smoothness in each direction (Stein, 2012).

6. Relations, distinctions, and recurrent themes

The supplied literature does not present a single universal theory of geometric screening. Instead, it presents several domain-specific theories that share structural motifs. One motif is projection onto constrained geometry: diagnostic curves are constrained to a unit square with invariant points; dual optimal points are projections onto convex polytopes; protein–ligand complexes are refined through equivariant coordinate updates in Euclidean space. A second motif is screening by induced structure: false positives and true positives reshape predictive-value curves, plastic quadrupoles generate dipoles and screening lengths, and host nanotube geometry alters induced dipoles. A third motif is safe exclusion by geometric bounds: optimization screening discards features, groups, or samples when a safe region excludes dual activity (Balayla, 2020).

Important distinctions remain. In optimization, “screening” generally means certifying zero contribution at the optimum, with safety guarantees and no loss of exactness (Wang et al., 2012). In drug discovery, screening is a ranking or enrichment task, and the geometric component resides in representation learning and coordinate generation rather than in exclusion certificates (He et al., 2024). In mechanics and nanostructures, screening denotes attenuation or amplification of fields, often quantified by an emergent length scale or by the ratio $\phi\in[0,1]$16 (Silvestrelli et al., 2024). In spatial prediction, it denotes asymptotic irrelevance of far observations under specified spectral and geometric conditions (Stein, 2012).

This suggests a minimal cross-domain characterization: geometric screening is the study of how geometry constrains, mediates, or certifies screening behavior. In some fields that geometry is Euclidean and constructive; in others it is variational, spectral, or topological. What unifies the cases in the supplied sources is not a single formula, but the insistence that screening becomes technically tractable when encoded in geometric objects such as curves, angles, convex safe sets, multipole fields, or screening lengths.