🧬 biology

TACTIVS: Test-time actives induction for flexible target-specific virtual screening

TACTIVS is a test-time active induction framework that leverages a pretrained molecular foundation model to perform flexible, target-specific virtual screening by combining label-free calibration and connectivity-weighted graph propagation, achieving competitive enrichment without requiring model fine-tuning even when query and support actives belong to distinct chemical series.

Original authors: Douglas Houston, Lin Chen, Haoying Wang, Andreas Luttens, Pedro Ballester

Published 2026-10-09

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Douglas Houston, Lin Chen, Haoying Wang, Andreas Luttens, Pedro Ballester

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ⚕️ This is an AI-generated explanation of a preprint that has not been peer-reviewed. It is not medical advice. Do not make health decisions based on this content. Read full disclaimer

In the early stages of discovering a new medicine, scientists face a needle-in-a-haystack problem. They must sift through vast libraries containing millions of chemical compounds to find the handful that might successfully bind to a specific disease target and stop it from causing harm. Traditionally, researchers have relied on two main strategies to narrow this search. One approach looks at the physical shape of the target, like a lock, and tries to find a key that fits. The other looks at known working keys—molecules already proven to activate the target—and searches for others that look similar. While the first method requires detailed 3D maps of the target, the second relies on having at least one working example to start with. The challenge arises when scientists have very few examples, perhaps only a single confirmed active molecule, and need to find more without wasting time on millions of dead ends.

A team of researchers at the University of Edinburgh, Karolinska Institutet, and Imperial College London has developed a new method called TACTIVS to solve this specific problem. Instead of trying to retrain a complex computer model every time a new active molecule is discovered, which usually requires a large amount of data, TACTIVS works by instantly reorganizing how the computer views the entire library of chemicals. The system takes a frozen, pre-trained understanding of molecular shapes and uses the few known active molecules to immediately reshape the search. It does this by creating a map where the known active molecules act as beacons. The system then traces paths through the library, not just looking for molecules that look exactly like the beacons, but also for those that are connected to them through a chain of chemical similarities. This allows the system to find promising candidates that might look quite different from the starting examples but still belong to the same functional family.

The researchers tested this approach on several large, standard datasets used to benchmark virtual screening performance. In one test involving nearly 150 different biological targets, the new method was able to identify active molecules far more effectively than existing techniques when given just a single known active molecule. When the team increased the number of known active examples to ten, the method's ability to find the correct compounds improved significantly, outperforming both traditional shape-matching tools and other advanced learning-based systems. Crucially, the method proved it could find active molecules even when they were chemically distinct from the starting examples. In a strict test where the known active molecules and the target compounds came from completely different chemical families, the system still managed to rank the correct compounds much higher than random chance would allow. This suggests the method is not merely finding close copies of the starting molecules but is successfully navigating the broader chemical landscape to find new types of effective drugs.

The core of this success lies in how the system handles the data. Rather than adjusting the internal settings of the computer model to fit the new target, which can be unstable with very little data, the researchers adjusted the way the data is measured. They applied a mathematical transformation that removes background noise and highlights the specific differences that matter for the current search. This allows the system to see the relationships between molecules more clearly. By combining a direct comparison with the known active molecules and a broader analysis of how molecules are connected to one another in the library, the system creates a ranking that evolves as more information becomes available. If a scientist discovers a new active compound during an experiment, they can simply add it to the list, and the system instantly recalculates the rankings without needing to be retrained.

This work demonstrates that the value of a single experimental result can be maximized by using it to guide a flexible, library-wide search. The findings suggest that for many drug discovery campaigns, especially those where data is scarce, the most effective strategy is not to build a new model for every target, but to use a shared, robust understanding of chemistry and adapt the search strategy on the fly. The researchers found that as more active molecules were added to the reference set, the quality of the search improved steadily, confirming that the method learns from the accumulating evidence. By proving that a system can remain fixed while the search strategy adapts to new targets, this approach offers a practical and efficient path forward for identifying new medicines in the early, data-poor stages of discovery.

Technical Summary: TACTIVS

Problem Statement
Virtual screening (VS) is a cornerstone of early-stage drug discovery, yet extracting maximum value from limited experimental data remains a significant challenge, particularly in the "few-shot" regime where only a small number of known active molecules are available. Traditional ligand-based virtual screening (LBVS) relies on predefined similarity metrics or alignment rules, which often fail to account for the broader organization of the candidate library. Conversely, recent structure-based few-shot methods frequently require additional target-specific training or model adaptation (e.g., fine-tuning parameters), which is computationally expensive and prone to overfitting when data is scarce. A central unresolved challenge is how to achieve flexible, target-specific rankings directly at test time without retraining encoders or fitting target-specific readouts, while still leveraging the relational structure of the entire screening library.

Methodology
The authors introduce TACTIVS (Test-time ACTIves induction for flexible target-specific VS), a ligand-based transductive inference framework. Unlike conventional approaches that treat candidates as independent similarity comparisons or require parameter updates, TACTIVS constructs a target-specific ranking by propagating evidence from known actives through a neighborhood graph of the entire candidate library.

The workflow consists of three primary stages:

Frozen Representation and Embedding Cache:
TACTIVS utilizes a pretrained Equivariant Pretrained Transformer (EPT) model. The encoder is kept frozen throughout the process. Molecular conformers from the screening library and reference actives are encoded into 512-dimensional vectors. These vectors are cached to avoid repeated encoding. Crucially, the framework employs a shared covariance estimation derived from an independent, large-scale molecular corpus (PDBscreen training partition) rather than the specific target library. This prevents statistical instability when target libraries are small.
Label-Free Calibration (Whitening):
To address the anisotropy of pretrained representations, TACTIVS applies a whitening transformation. It centers the embeddings using the molecule-balanced mean of the current target library and applies a symmetric inverse-square-root whitener based on the shared, shrinkage-regularized covariance matrix. This transforms the embeddings into a space where directions of high variance in the pretraining distribution do not dominate the similarity metric, effectively making the information more accessible to simple similarity-based retrieval.
Transductive Inference via Graph Propagation:
The inference combines two sources of evidence:
- Direct Active-Reference Similarity: Candidates are scored based on their direct similarity to the conformers of the known active references.
- Graph-Propagated Relational Evidence: A weighted neighborhood graph is constructed over the entire library (including both actives and candidates) using the calibrated embeddings. Known actives "seed" the graph with activity evidence. This evidence is then propagated through the graph using a random walk with restart mechanism. This allows candidates that are not direct neighbors of an active to still receive support if they are connected via intermediate molecules, capturing multi-step molecular paths.
The final ranking score is a linear combination of the standardized direct similarity and the propagated graph score. The framework requires no encoder fine-tuning or target-specific readout fitting; it adapts solely through the selection of reference molecules and the library context.

Key Contributions

Test-Time Adaptation without Training: TACTIVS achieves target specificity by contextualizing known actives within the candidate manifold at inference time, eliminating the need for parameter updates or additional training data.
Transductive Graph Propagation: It introduces a library-level inference approach where the screening collection itself becomes part of the inference process, allowing evidence to flow through the molecular network rather than relying solely on pairwise similarity.
Robust Covariance Calibration: The use of a shared, external covariance estimate for whitening ensures stable calibration even when the target library is small or chemically redundant, avoiding the instability of estimating covariance from few-shot data.
Chemical Series Generalization: The method is designed to retrieve active molecules that belong to distinct chemical series from the reference set, moving beyond the retrieval of close structural analogues.

Results
The authors evaluated TACTIVS on two major benchmark families: VSDS-vd (TrueDecoy and RandomDecoy subsets) and LIT-PCBA.

Performance on VSDS-vd: TACTIVS significantly outperformed state-of-the-art baselines, including structure-based docking (e.g., Glide, AutoDock), hybrid methods, and other ligand-based approaches. With a single active reference, TACTIVS achieved an EF@0.5% of 19.09 on TrueDecoy, substantially exceeding the best published baseline of 6.91. Performance improved consistently as the number of reference actives ( $K$ ) increased, reaching an EF@1% of 34.13 at $K=10$ .
Performance on LIT-PCBA: In a comparison against zero-shot models (ConGLUDe) and few-shot adaptation models (Drug-few), TACTIVS demonstrated superior early enrichment. At $K=8$ , it achieved a BEDROC of 0.182, surpassing both the 8-shot Drug-few model and the zero-shot ConGLUDe model.
Series-Disjoint Evaluation: Under stringent evaluation protocols where reference and query actives belonged to distinct chemical series (ECFP4 Tanimoto similarity < 0.5), TACTIVS retained substantial early enrichment. This confirmed that the method does not merely recover close analogues but can generalize to novel chemotypes within the library.
Ablation Studies: Removing the covariance whitening transformation resulted in a significant drop in performance (e.g., a 20.32 point drop in EF@1% on RandomDecoy), highlighting the importance of the calibration step. Similarly, removing the graph propagation component reduced performance, confirming the value of relational evidence.

Significance and Claims
The paper claims that TACTIVS establishes pretrained molecular representations as adaptable substrates for flexible target-specific screening. The authors argue that target specificity does not necessarily require parameter updates; instead, it can emerge naturally when a general inference procedure contextualizes known actives within the candidate manifold.

The work proposes a shift in perspective for virtual screening: distinguishing between de novo screening (where no actives are known) and incremental screening (where experimental hits are available). TACTIVS addresses the latter, providing a practical route to maximizing the value of scarce experimental data by allowing new actives to inform subsequent rankings immediately without retraining. The authors position this as a form of generalization where a shared representation and inference rule are specialized at test time through transductive reasoning, offering a data-efficient alternative to few-shot learning paradigms that rely on parameter adaptation.

Technical Summary: TACTIVS

More like this