Transferability Index in Computational Dynamics

• Transferability Index (TI) is a metric that quantifies how well models, features, or representations from one system can be applied to another.
• It evaluates cross-domain transfer by comparing statistical descriptors, such as time-lagged independent components, to assess dynamic similarity.
• TI guides enhanced sampling by identifying when transferring collective variables can reduce computational cost and improve simulation accuracy.

The Transferability Index (TI) quantifies how well knowledge, models, or representations obtained from one domain, system, or task (the "source") can be applied to another (the "target"). In the context of computational science and machine learning, TI formalizes the notion of transferability—the potential for effective transfer learning or reuse—by providing a systematic metric or set of metrics that estimate the performance, utility, or suitability of cross-domain, cross-task, or cross-system transfers. Accurate transferability estimation is especially critical in high-dimensional or computationally expensive domains, such as molecular dynamics, computer vision, and LLMing, where exhaustive retraining or resimulation is often infeasible.

1. Methodological Foundations of the Transferability Index

The TI concept rests on comparing key properties of knowledge or models extracted from source and target settings. In atomistic molecular simulation, this specifically refers to time-lagged independent components (TICs), statistical descriptors capturing the slowest modes of motion, which are typically obtained using time-lagged Independent Component Analysis (TICA). The transferability of these TICs can serve as a practical proxy for broader transferability questions in dynamical systems.

Formally, for an $N$-dimensional vector time series $\mathbf{r}_t \in \mathbb{R}^N$, TICA solves: $$\mathbf{C}^r(\tau)\mathbf{U} = \bm{\Lambda}\mathbf{C}^r(0)\mathbf{U}$$ where $\mathbf{C}^r(0)$ and $\mathbf{C}^r(\tau)$ are the covariance and time-lagged covariance matrices, $\mathbf{U}$ is the matrix of TICs, and $\bm{\Lambda}$ is a diagonal matrix of autocorrelation eigenvalues.

The TI comprises a set of quantifiable, interpretable metrics that assess how well TICs determined on a donor system describe the slow dynamics of an acceptor system. The TI thus enables rigorous, interpretable, system-agnostic evaluation of collective variable transferability.

2. Definitions and Key Quantities

The paper defines three core metrics for evaluating TIC transferability:

1. Covariance Similarity ($D_0$):

$$D_0 \equiv \left\| \mathbf{C}^v(0) - \mathbf{I} \right\|_F$$

Measures the deviation of the covariance matrix of donor TICs projected on the acceptor from the identity, with $$\mathbf{C}^v(0) = \mathbf{V}^\top \mathbf{C}^r(0)\mathbf{V}$$.

1. Time-lagged Covariance Similarity ($D_\tau$):

$$D_\tau \equiv \left\| \mathbf{C}^v(\tau) - \bm{\Lambda} \right\|_F$$

Quantifies the consistency between the time-lagged covariance of projected donor TICs and the acceptor's native TIC autocorrelations.

1. Subspace Overlap ($D_{KM}$):

$$D_{KM} \equiv \left\| \mathbf{V}_K \mathbf{X} - \mathbf{U}_M \right\|_F$$

Where $\mathbf{X} = \mathbf{V}_K^+ \mathbf{U}_M$ (the Moore-Penrose pseudoinverse), $D_{KM}$ characterizes how well the leading $K$ donor TICs span the subspace of the top $M$ acceptor TICs.

Lower values of each metric reflect greater transferability; a value of zero signals perfect transfer.

3. Workflow and Implementation

A standard protocol for applying TI to evaluate transferability in practical settings includes:

1. Compute TICs for both donor and acceptor systems via TICA, obtaining $\mathbf{U}$ (acceptor) and $\mathbf{V}$ (donor).
2. Project donor TICs onto acceptor data.
3. Calculate $D_0$, $D_\tau$, and $D_{KM}$ using the definitions above.
4. Assess transfer efficiency via the "relative transfer time," defined as the fraction of acceptor system sampling needed for native TIC performance to match that achievable via transferred donor TICs.
5. Interpret results in the context of system-specific factors, correlating with structural, kinetic, and thermodynamic similarities when possible.

A summarized implementation flow:

U = TICA(acceptor_data) V = TICA(donor_data) C0 = covariance_matrix(acceptor_data) Ctau = time_lagged_covariance_matrix(acceptor_data, tau) D0 = frobenius_norm(V.T @ C0 @ V - np.eye(K)) Dtau = frobenius_norm(V.T @ Ctau @ V - Lambda) X = np.linalg.pinv(V_K) @ U_M DKM = frobenius_norm(V_K @ X - U_M)

4. System Dependence and Empirical Observations

TI values are highly sensitive to the type and magnitude of system perturbation:

• Symmetry: Transferability is often, but not always, bidirectional. Some protein mutants (e.g., Met-enkephalin $\rightarrow$ Leu-enkephalin and vice versa) exhibit nearly symmetric transfer, while others demonstrate strong asymmetries.
• Perturbation magnitude: Small modifications that preserve major metastable states and their connectivity (e.g., sidechain mutations) tend to result in higher TI, but this is not guaranteed.
• Landscape changes: Addition of new metastable wells often preserves transferability (high TI), whereas the deletion of important states may irreversibly destroy slow modes and drastically reduce TI, regardless of overall structural similarity.
• Noise and convergence: Finite sampling and poor exploration can artificially inflate TI metrics due to sampling noise.

A representative table of empirical relative transfer times:

Donor System	Acceptor System	Relative Transfer Time
Met-enkephalin	Leu-enkephalin	0.170
Leu-enkephalin	Met-enkephalin	0.154
GTT mutant	FiP35 WW Domain	0.171
Apo Calmodulin	Holo Calmodulin	0.002

5. Factors Affecting Transferability

Transferability, as quantified by TI, is dictated by:

• Degree of shared metastable structure: Systems with similar free energy surfaces along slow TICs tend to achieve higher TI, but exceptions are common.
• Dynamics and pathways: Changes impacting slow dynamical processes can decrease transferability even when equilibrium distributions appear similar.
• Sampling sufficiency: Proper convergence of the donor and acceptor TICA models is critical for meaningful TI assessment.

The use of toy models in the paper demonstrates that changes in the topology of the potential energy landscape (additions or deletions of basins) can be detected as abrupt changes in TI, providing mechanistic insight.

6. Applications and Implications

The TI is directly applicable to:

• Enhanced sampling design: Justifies using transferred TICs as collective variables for accelerated simulation only when TI is high, minimizing wasted computational effort.
• Perturbation Impact Analysis: Enables rigorous quantification of how mutations, ligand binding, or environmental changes alter slow system kinetics, extending beyond static characterization.
• Framework for comparative analysis: Offers a quantitative, interpretable method for evaluating the dynamical similarity of molecular systems, complementary to clustering or kinetic model-based methods.

In enhanced sampling (e.g., TICA-metadynamics), a high TI encourages transferring collective variables, leading to improved convergence, while a low TI signals the need for either new simulation or hybrid modeling.

---

The Transferability Index (TI), as realized through rigorous distance metrics comparing TIC properties, provides a foundational, robust, and interpretable approach for assessing when, and to what extent, information learned in one system can be leveraged in another. TI supports reproducible, quantitative model and variable selection in computational molecular science but is system-specific: high TI reliably predicts successful transfer, while low TI flags the risk of inadequate or misleading transferral.