Benchmarking Large Language Models for Assay-Aware Drug-Target Binding Affinity Prediction

Gupta, N.; Verma, A.; Khandelwal, P.; Sethi, T.

doi:10.64898/2026.10.11.758217

Home/Archive/bioRxiv

bioRxiv · 10.64898/2026.10.11.758217

Benchmarking Large Language Models for Assay-Aware Drug-Target Binding Affinity Prediction

Gupta, N.·Verma, A.·Khandelwal, P.·Sethi, T.

View original source ↗Export metadata ↓

Abstract

Machine learning models for molecular discovery often train on bioactivity data from heterogeneous assays, yet assay context is rarely explicitly represented. We investigate whether natural-language assay descriptions provide predictive information beyond molecular and protein representations. We develop ACL-DTI (Assay-aware Contrastive Learning-based Drug Target Interaction) and MAMMAL-AsAP, an assay-context extension of MAMMAL-DTI. Across five triplet-disjoint random splits, assay-aware models achieve competitive or improved performance over assay-agnostic baselines. Among the evaluated models, MAMMAL-AsAP achieves the highest (R^2) of 0.7198 +/- 0.0085. Masking experiments show that ACL-DTI is robust to incomplete assay descriptions, exhibiting the smallest reductions in R^2 among the evaluated models when masking the target name, assay technique, or both: 0.0228 +/- 0.0064, 0.0406 +/- 0.0113, and 0.0620 +/- 0.0190, respectively. These results indicate that assay descriptions contain predictive information not fully captured by molecular structures and protein sequences, and reveal architecture-dependent differences in how this information is exploited.

Explore related subjects

bioinformatics Preprint

Keep this discovery

Explore connections, maps & timelines

BibTeX RIS

Gupta, N., Verma, A., Khandelwal, P., Sethi, T.. 2026-10-16. Benchmarking Large Language Models for Assay-Aware Drug-Target Binding Affinity Prediction. https://doi.org/10.64898/2026.10.11.758217

Cite the original work for its findings. Save a collection to share your selection of sources.

KEEP EXPLORING

Related preprints

bioRxiv2026

Phase-Based Frequency Tracking and Analysis of Vocalisations in the Genus Saxicola

Acoustic analysis provides valuable tools for studying bird vocalisations, but quantitative characterisation of call structure remains challenging when frequency and amplitude vary jointly over time. We present a phase-based frequency tracking method that represents amplitude as a function of instantaneous frequency and use it to characterise alarm calls in the genus Saxicola. The method enables the extraction of acoustic features describing the frequency trajectory and temporal structure of individual calls. We apply the method to a dataset of 801 calls from 49 recordings representing 15 species, and use principal component analysis (PCA) and robust PCA to investigate patterns of acoustic variation across species, recordings and geographic regions. The results demonstrate the feasibility of the approach and reveal structured patterns of acoustic similarity and variation among the calls analysed. Robust PCA concentrates the variance of the low-rank representation into fewer principal components than conventional PCA applied to the original feature matrix, providing a complementary representation of the dominant acoustic structure. The observed patterns illustrate the potential of quantitative acoustic characterisation for comparative studies of vocal variation. However, uneven species representation and uncertainty in individual identity and taxonomic labels limit the biological conclusions that can be drawn from the present dataset. The method provides a foundation for further acoustic research, in which broader sampling and integration with independent genetic and morphological evidence could help investigate variation within and among species.

bioinformatics↗

bioRxiv2026

A Sparse Modal Representation for Dolphin Echolocation Transients

Characterising echolocation provides a non-invasive means of monitoring cetacean populations, health, and behaviour, offering critical data to inform conservation policy and mitigate the impacts of anthropogenic ocean noise. Echolocation clicks are short broadband transients whose internal oscillatory structure is difficult to compare in a sequence using conventional time- or frequency-domain representations. We introduce a Sparse Modal Representation (SMR) in which each click is described as a sparse superposition of delayed damped resonant modes. The representation separates resonant behaviour from temporal localisation, allowing both narrowband and broadband click components to emerge from the same set of resonant building blocks. The methodology combines Burg autoregressive pole estimation, sparse approximation in an over-complete dictionary, and nonlinear parameter refinement to obtain compact and interpretable modal descriptions of individual transients. The method is evaluated using recordings of free-ranging common dolphins (Delphinus delphis) from the DOLPHINFREE project in the Northern Bay of Biscay. The results demonstrate accurate reconstruction using a limited set of parameters that are repeatable across successive clicks and robust under realistic recording conditions. These conditions include wave, vessel, and sonar interference. Repeatability and robustness are essential for using this parameter set to detect changes in animal behaviour.

bioinformatics↗

bioRxiv2026

muDM: A Queryable Multi-Resolution Data Model for Vector Geometry in Spatial Bioimaging

Spatial bioimaging combines raster imagery with vector geometries that encode molecular measurements, connectivity, and anatomical labels. A major challenge is to connect and jointly query these resources at scale, because the vector structures that locate them cannot be searched across resolutions. muDM is a spatial data model for associating vector geometry with its annotated records. Under a single schema, it arranges 2D and 3D geometries, together with their identities and measurements, as a precomputed multi-resolution pyramid, a columnar store that can be queried in place from object storage and equivalent render-ready tiles from the same index. Deployed in a public catalog containing more than 100 datasets, this single model supports connectome meshes, multiplexed tissue imaging, and spatial transcriptomics. For neuron meshes, classifiers which use geometry together with linked synapse counts exceed the performance of geometry-only models, while coarser levels preserved most balanced accuracy and using up to two orders of magnitude less stored geometry.

bioinformatics↗