Affinity Tailor: Structuring Relational Metrics

• Affinity Tailor is a cross-domain methodological pattern that explicitly structures affinity through geometry, topology, and attribute interactions.
• It leverages domain-specific techniques—such as latent-space metric learning for drug–target binding and graph-based lineage analysis in antibody prediction—to align affinity with task structure.
• Empirical results demonstrate enhanced performance and interpretability by tailoring relational metrics rather than relying on undifferentiated similarity measures.

“Affinity Tailor” denotes a family of methods that explicitly shape, condition, or exploit an affinity structure rather than treating compatibility, similarity, or driving force as an undifferentiated scalar. In the supplied literature, the term appears in several technically distinct settings: drug–target binding prediction, social affinity modeling from personality and language, affinity-aware scheduling in multicore systems, antibody and aptamer affinity analysis, multi-target tracking, complementary fashion recommendation, and generalized thermodynamic force definitions. Across these settings, the common motif is the same: affinity is not merely measured after representation learning or system evolution, but is actively structured through geometry, locality, attribute interactions, lineage information, or generalized entropy formalism (Refahi et al., 25 Sep 2025, Tshimula et al., 2022, Ng et al., 30 Apr 2026, Rodríguez-Castellanos et al., 12 Jun 2026).

1. Conceptual scope and recurring design pattern

In drug–target affinity prediction, “Affinity Tailor” is presented as a geometry-aware, alignment-based paradigm in which ligand and protein embeddings are conditioned and organized so that binding strength corresponds to distance in latent space (Refahi et al., 25 Sep 2025). In personality and social network analysis, the term denotes an end-to-end framework that derives affinity relationships from social media interactions, predicts personality from language, and analyzes semantic similarity and emotional stability across Myers–Briggs personality types (Tshimula et al., 2022). In scheduling, “Affinity Tailor” is the name of a production system that assigns demand-sized, topologically compact preferred CPU sets and treats them as soft affinity hints rather than hard partitions (Ng et al., 30 Apr 2026).

Other usages follow the same structural idea. In complementary fashion recommendation, AFRec quantifies the affinity of a pair of items through learned compatibility between their explicit attributes, with attributes serving as the bridge between two fashion items (Li et al., 2021). In multi-target multi-camera tracking, an adaptive affinity module tailors the metric to the local matching scope of data association rather than using global re-identification distances unchanged (Hou et al., 2021). In antibody repertoire analysis, affinity prediction is approached through lineage-derived sequence and branching metrics such as amino-acid consensus distance and local branching ratio (Ralph et al., 2020). In aptamer–ligand analysis, affinity is probed through graph hierarchy, assortativity, and electrical resistance (Cataldo et al., 2017). In thermodynamics, the notion is generalized into $q$-affinities associated with Rényi and Tsallis entropies (Rodríguez-Castellanos et al., 12 Jun 2026).

This suggests that “Affinity Tailor” is best understood not as a single algorithm but as a cross-domain methodological pattern: affinity is made explicit, structured, and domain-aligned, rather than left implicit inside a generic predictor or system controller.

2. Geometry-aware and latent-space formulations

A prominent technical instantiation appears in "Learning to Align Molecules and Proteins: A Geometry-Aware Approach to Binding Affinity" (Refahi et al., 25 Sep 2025). The model FIRM-DTI defines drug–target affinity prediction as learning a function

$$f:\ (\text{ligand}, \text{protein}) \mapsto \text{binding affinity},$$

with geometry referring to the geometry of the learned embedding space rather than explicit 3D coordinates. Each drug and protein is mapped to a vector in $\mathbb{R}^d$, and distances are constrained so that high-affinity pairs lie close while low-affinity or non-interacting pairs lie far apart (Refahi et al., 25 Sep 2025).

The architecture combines separate encoders, FiLM conditioning of ligand by protein, normalized cosine distance, and an RBF regression head. The conditioning step is

$$\gamma(z_t) = W_\gamma z_t + b_\gamma,\quad \beta(z_t) = W_\beta z_t + b_\beta,$$

$$\tilde{z}_d = \mathrm{FiLM}(z_d \mid z_t) = \gamma(z_t) \odot z_d + \beta(z_t),$$

followed by normalized embeddings and cosine distance

$$d = 1 - \hat{z}_d \cdot \hat{z}_t,\quad d \in [0,2].$$

Affinity is then predicted as a smooth univariate function of distance via radial basis functions (Refahi et al., 25 Sep 2025). On the Therapeutics Data Commons DTI-DG benchmark, the reported test performance is PCC = 0.59, with ablations showing 0.55 without FiLM and 0.32 without triplet loss, indicating that conditioning and metric learning are central rather than cosmetic (Refahi et al., 25 Sep 2025).

A related latent-space formulation appears in "Co-Diffusion: An Affinity-Aware Two-Stage Latent Diffusion Framework for Generalizable Drug-Target Affinity Prediction" (Qian et al., 11 Mar 2026). There, Stage I learns an affinity-steered latent manifold under direct supervised regression, while Stage II adds modality-specific latent diffusion as a perturb-and-denoise regularizer. The framework treats drug and target latent variables as clean semantic latents,

$$z_{d,0} = \mu_d + \sigma_d \odot \varepsilon_d,\quad z_{t,0} = \mu_t + \sigma_t \odot \varepsilon_t,$$

then diffuses and denoises them while preserving affinity semantics through a second regression head (Qian et al., 11 Mar 2026). The paper explicitly frames the design as an “Affinity Tailor” that sculpts an affinity-aware latent manifold and then regularizes it with latent diffusion (Qian et al., 11 Mar 2026).

These formulations share a strong inductive bias: affinity is encoded as geometry in a representation space, and robustness is enforced by metric structure or denoising consistency rather than by simple concatenation of modality features. A plausible implication is that “Affinity Tailor” in this line of work refers to the deliberate shaping of latent manifolds so that affinity becomes a first-class coordinate of representation.

3. Graph, lineage, and metric tailoring

A second major family of “Affinity Tailor” methods relies on explicit graph or lineage structure. In aptamer–ligand modeling, "Hierarchy and assortativity as new tools for affinity investigation: the case of the TBA aptamer-ligand complex" represents the thrombin-binding aptamer and the TBA–thrombin complex as graphs whose nodes are nucleobases and amino acids, with edges defined by spatial proximity under a cut-off radius $R_c$ (Cataldo et al., 2017). The analysis uses degree distribution, rank–degree hierarchy

$k(r) = A e^{a r},$

and assortativity via

$\langle k_{nn} \rangle(k) = D + b k,$

to distinguish Na$$f:\ (\text{ligand}, \text{protein}) \mapsto \text{binding affinity},$$0 and K$$f:\ (\text{ligand}, \text{protein}) \mapsto \text{binding affinity},$$1 complexes (Cataldo et al., 2017). The paper interprets a transition from disassortative “open” networks in TBA alone to assortative “closed” networks in the TBA–thrombin complex as a topological signature of high affinity, and also maps the graph to an electrical network where resistance becomes an affinity-sensitive physical descriptor (Cataldo et al., 2017).

In antibody affinity prediction, "Using B cell receptor lineage structures to predict affinity" develops sequence-level and branch-level scores derived from clonal family structure (Ralph et al., 2020). The most effective sequence-level score is amino-acid consensus distance, aa-cdist, defined as the amino-acid Hamming distance to the family consensus sequence. Tree-based measures include local branching index, $$f:\ (\text{ligand}, \text{protein}) \mapsto \text{binding affinity},$$2, and local branching ratio, $$f:\ (\text{ligand}, \text{protein}) \mapsto \text{binding affinity},$$3, where the latter compares branch length below a node to branch length above and beside it to identify likely affinity-increasing mutations (Ralph et al., 2020). The authors report that aa-cdist is the best single predictor of affinity across wide simulation regimes, while lbr localizes affinity-increasing mutations to within approximately $$f:\ (\text{ligand}, \text{protein}) \mapsto \text{binding affinity},$$4–$$f:\ (\text{ligand}, \text{protein}) \mapsto \text{binding affinity},$$5 ancestors from perfect in simulation (Ralph et al., 2020).

In multi-target multi-camera tracking, "Adaptive Affinity for Associations in Multi-Target Multi-Camera Tracking" argues that global re-identification feature distances are mismatched to local association problems, because re-ID is global while tracking is local in space and time (Hou et al., 2021). The method replaces a global distance-derived affinity

$$f:\ (\text{ligand}, \text{protein}) \mapsto \text{binding affinity},$$6

with a learned metric based on the absolute feature difference,

$$f:\ (\text{ligand}, \text{protein}) \mapsto \text{binding affinity},$$7

and outputs

$$f:\ (\text{ligand}, \text{protein}) \mapsto \text{binding affinity},$$8

The critical innovation is temporal-window sampling aligned to the actual association windows used in single-camera and multi-camera tracking (Hou et al., 2021). On CityFlow, the adaptive affinity configuration reports IDF1 = 63.0 versus 56.6 for re-ID distance, and on DukeMTMC it improves MCT IDF1 on the hard test set from 75.4 to 82.3 (Hou et al., 2021).

These works share a consistent methodological move: affinity is tailored to the operative graph neighborhood, lineage neighborhood, or association window, rather than assumed to be globally homogeneous.

4. Attribute-, personality-, and language-mediated affinity

In social affinity analysis, "Discovering Affinity Relationships between Personality Types" operationalizes affinity relationships as friendship-like ties characterized by mutual understanding, reciprocal and common interests, sympathy, harmonious communication, and agreement between individuals, inferred from social media interactions using HAR-search (Tshimula et al., 2022). The resulting affinity graph is

$$f:\ (\text{ligand}, \text{protein}) \mapsto \text{binding affinity},$$9

with weighted edges $\mathbb{R}^d$0, and the personality-annotated graph is

$\mathbb{R}^d$1

with label mapping $\mathbb{R}^d$2 over the 16 MBTI types (Tshimula et al., 2022).

The framework combines several affinity-related layers. Personality is predicted from spontaneous language using classifiers including Logistic Regression, Random Forest with AdaBoost, SVM, Naive Bayes, and BERT-large-cased, with the paper reporting F1 > 0.8 for many types under BERT and stating that personality can be predicted from spontaneous language with an F-1 score superior to 0.76 (Tshimula et al., 2022). Type-level semantic similarity is quantified by cosine similarity between GloVe-based type corpora, and emotional stability is assessed by Pearson correlations on LIWC positive and negative emotion features (Tshimula et al., 2022). Clustering with MCL and K-destinations identifies ESTJ, INTP, ENFP, ENTJ as robustly influential personality types in affinity formation (Tshimula et al., 2022).

In complementary fashion recommendation, AFRec explicitly models compatibility through attribute-level affinity (Li et al., 2021). Each item is represented by a matrix of attribute embeddings,

$\mathbb{R}^d$3

and reciprocal attention yields attribute-importance vectors for the top and bottom conditioned on each other (Li et al., 2021). The method computes a category-specific bilinear compatibility matrix,

$\mathbb{R}^d$4

an explicit affinity matrix,

$\mathbb{R}^d$5

and a weighted compatibility matrix

$\mathbb{R}^d$6

whose sum gives the final compatibility score (Li et al., 2021). On FashionVC and PolyvoreMaryland, AFRec reports the best AUC and HR@40 among the listed baselines, including AUC 0.741 / HR@40 0.789 on FashionVC and AUC 0.753 / HR@40 0.828 on PolyvoreMaryland (Li et al., 2021).

These studies illustrate a different but related meaning of “Affinity Tailor”: explicit semantic mediators—personality types, linguistic signals, or item attributes—are inserted into the scoring process so that affinity is both more discriminative and more interpretable.

5. Systems and hardware interpretations of affinity

The term also appears in systems work, most directly in "Affinity Tailor: Dynamic Locality-Aware Scheduling at Scale" (Ng et al., 30 Apr 2026). Here affinity does not refer to molecular binding or social compatibility, but to CPU locality: workloads preserve locality when their threads are steered toward compact, demand-sized CPU sets, especially on chiplet-based systems with multiple LLC domains (Ng et al., 30 Apr 2026). The architecture has a userspace controller in Borglet that estimates each container’s CPU demand from 1-second utilization samples over a 5-minute window and defines

$\mathbb{R}^d$7

typically with $\mathbb{R}^d$8 (Ng et al., 30 Apr 2026). The kernel then treats per-cgroup preferred CPU sets as soft affinity hints rather than hard partitions (Ng et al., 30 Apr 2026).

The reported production results are concrete: geometric-mean per-CPU throughput gains of 12% on chiplet-based systems and 3% on non-chiplet systems over Linux CFS, with per-GB throughput gains of 3–7% (Ng et al., 30 Apr 2026). The paper also reports improved Preferred Core Residency, lower branch MPKI, and large reductions in LLC miss rate on split-LLC platforms, while accepting a P99 scheduling-latency increase of up to approximately 17% (Ng et al., 30 Apr 2026). The system’s central claim is that future schedulers should treat spatial locality as a first-class objective, even at the expense of work-conservation (Ng et al., 30 Apr 2026).

A different systems usage appears in "Tailor: Altering Skip Connections for Resource-Efficient Inference" (Weng et al., 2023). Although not titled “Affinity Tailor,” it belongs to the same naming family and treats skip-connection structure as something that can be tailored to hardware constraints after training (Weng et al., 2023). Tailor uses a frozen teacher, a trainable student, and a knowledge-distillation loss

$\mathbb{R}^d$9

then progressively removes or shortens skip connections during retraining (Weng et al., 2023). The reported hardware effects include up to 34% BRAM, 13% FF, and 16% LUT savings for hls4ml-style FPGA designs, plus about 30% throughput improvement and 45% memory-bandwidth reduction for a 2D PE array on ResNet-50 (Weng et al., 2023).

These usages widen the meaning of “Affinity Tailor” beyond affinity as similarity. In systems settings, affinity becomes a placement or structural-bias relation—between workload and cores, or between learned network topology and hardware topology. This suggests a broader editorial reading: the term often signals deliberate co-design of a relational structure with its execution environment.

6. Thermodynamic and materials-science extensions

In "Thermodynamic Framework for $$\gamma(z_t) = W_\gamma z_t + b_\gamma,\quad \beta(z_t) = W_\beta z_t + b_\beta,$$0-Affinity," affinity is generalized from the classical De Donder driving force to non-equilibrium systems described by Rényi or Tsallis entropies (Rodríguez-Castellanos et al., 12 Jun 2026). Classical affinity is

$$\gamma(z_t) = W_\gamma z_t + b_\gamma,\quad \beta(z_t) = W_\beta z_t + b_\beta,$$1

with entropy production

$$\gamma(z_t) = W_\gamma z_t + b_\gamma,\quad \beta(z_t) = W_\beta z_t + b_\beta,$$2

The paper then defines generalized $$\gamma(z_t) = W_\gamma z_t + b_\gamma,\quad \beta(z_t) = W_\beta z_t + b_\beta,$$3-affinities. For Rényi entropy,

$$\gamma(z_t) = W_\gamma z_t + b_\gamma,\quad \beta(z_t) = W_\beta z_t + b_\beta,$$4

which can be written using the escort distribution as an escort-weighted average score (Rodríguez-Castellanos et al., 12 Jun 2026). A generalized potential

$$\gamma(z_t) = W_\gamma z_t + b_\gamma,\quad \beta(z_t) = W_\beta z_t + b_\beta,$$5

leads to

$$\gamma(z_t) = W_\gamma z_t + b_\gamma,\quad \beta(z_t) = W_\beta z_t + b_\beta,$$6

and the same framework connects $$\gamma(z_t) = W_\gamma z_t + b_\gamma,\quad \beta(z_t) = W_\beta z_t + b_\beta,$$7 to Jarzynski-type exponential work averages (Rodríguez-Castellanos et al., 12 Jun 2026). For Tsallis entropy and Markov jump processes, the paper derives a $$\gamma(z_t) = W_\gamma z_t + b_\gamma,\quad \beta(z_t) = W_\beta z_t + b_\beta,$$8-deformed entropy balance law and a local stochastic $$\gamma(z_t) = W_\gamma z_t + b_\gamma,\quad \beta(z_t) = W_\beta z_t + b_\beta,$$9-affinity,

$$\tilde{z}_d = \mathrm{FiLM}(z_d \mid z_t) = \gamma(z_t) \odot z_d + \beta(z_t),$$0

while proving non-negativity of the associated entropy production rate (Rodríguez-Castellanos et al., 12 Jun 2026).

In materials science, "Impact of surface treatments on the electron affinity of nitrogen-doped ultrananocrystalline diamond" uses “tailoring” in the literal sense of tuning electron affinity via surface chemistry (Chambers et al., 2024). Electron affinity is defined as

$$\tilde{z}_d = \mathrm{FiLM}(z_d \mid z_t) = \gamma(z_t) \odot z_d + \beta(z_t),$$1

and practically estimated as

$$\tilde{z}_d = \mathrm{FiLM}(z_d \mid z_t) = \gamma(z_t) \odot z_d + \beta(z_t),$$2

with $$\tilde{z}_d = \mathrm{FiLM}(z_d \mid z_t) = \gamma(z_t) \odot z_d + \beta(z_t),$$3 assumed for calculation (Chambers et al., 2024). The measured values for N-UNCD are all positive: approximately +0.4 eV for hydrogen plasma treated, +1.3 eV for as-grown, +2.3 eV for oxygen plasma treated, and +2.1 eV for oxygen annealed surfaces (Chambers et al., 2024). A central result is that hydrogen-terminated N-UNCD exhibits positive electron affinity, in contrast to the widely reported negative electron affinity of hydrogen-terminated single-crystal diamond, owing to increased surface and bulk defect densities (Chambers et al., 2024).

These domains show the semantic breadth of the term. In thermodynamics, affinity is a generalized force; in materials science, it is an electronic band-edge quantity; in both cases, “tailoring” denotes controlled deformation of that quantity through formalism or surface treatment.

7. Shared principles, limitations, and interpretive synthesis

Despite disciplinary differences, the supplied works converge on several recurrent principles. First, affinity is treated as structured rather than primitive. In FIRM-DTI, latent distance is forced to encode affinity through triplet loss and RBF regression (Refahi et al., 25 Sep 2025). In Co-Diffusion, a latent manifold is made affinity-aware before generative regularization is applied (Qian et al., 11 Mar 2026). In AFRec, item compatibility is decomposed into explicit attribute–attribute interactions (Li et al., 2021). In adaptive tracking, the affinity metric is retrained for the local temporal windows actually encountered at inference time (Hou et al., 2021). In Affinity Tailor scheduling, CPU affinity is treated as a soft, demand-sized, topology-aware preference rather than a binary partition (Ng et al., 30 Apr 2026).

Second, most instances reject globally uniform scoring. The social-affinity work distinguishes graph-level affinity, personality-type affinity frequencies, semantic similarity, and emotional stability rather than collapsing them into one measure (Tshimula et al., 2022). The antibody-lineage work shows that amino-acid consensus distance substantially outperforms a naive mutation-count heuristic, implying that evolutionary context matters more than raw divergence (Ralph et al., 2020). The aptamer study likewise argues that hierarchy, assortativity, and resistance reveal affinity-relevant structure not visible in degree counts alone (Cataldo et al., 2017).

Third, interpretability is often strengthened by the tailoring step itself. AFRec’s weighted compatibility matrix is directly part of the score computation, not a post hoc explanation (Li et al., 2021). FIRM-DTI produces a one-dimensional distance-to-affinity curve with reported $$\tilde{z}_d = \mathrm{FiLM}(z_d \mid z_t) = \gamma(z_t) \odot z_d + \beta(z_t),$$4 correlation between distance and the observed affinity curve (Refahi et al., 25 Sep 2025). The social-affinity framework exposes type-pair frequencies, emotional correlations, and clustering-defined influential types (Tshimula et al., 2022). The thermodynamic framework preserves a force–flux entropy-production structure even under generalized entropy (Rodríguez-Castellanos et al., 12 Jun 2026).

The limitations are equally consistent. Many methods depend on strong upstream representations or pretrained experts: MolE and ESM2 in FIRM-DTI (Refahi et al., 25 Sep 2025), BERT-large-cased in personality prediction (Tshimula et al., 2022), and reliable re-ID embeddings in adaptive tracking (Hou et al., 2021). Several omit explicit 3D structure or fine-grained uncertainty, as noted for FIRM-DTI and Co-Diffusion (Refahi et al., 25 Sep 2025, Qian et al., 11 Mar 2026). Others rely on inferred labels or graph structure that may be noisy, such as self-reported MBTI types (Tshimula et al., 2022) or inferred BCR lineage trees (Ralph et al., 2020). In systems settings, locality-aware tailoring may increase scheduling latency or depend on non-pathological overcommitment and high-quality topology information (Ng et al., 30 Apr 2026).

Taken together, these works support a broad technical definition: an “Affinity Tailor” is a method that engineers the representation, locality, topology, or force law governing affinity so that the operative notion of closeness, compatibility, or driving tendency matches the real structure of the task. This suggests that the term’s unifying significance lies less in any single formalism than in a shared methodological stance: affinity should be designed, not merely measured.