Generative Retrieval for Reverse Target Fishing with a Proteome-Scale Benchmark

Zhichao Deng; Haichuan Tan; Xin Hong; Huaizhi Wang; Tao Huang; Bowen Gao; Yanyan Lan

cs.LGOct 17, 2026

Generative Retrieval for Reverse Target Fishing with a Proteome-Scale Benchmark

Authors: Zhichao Deng, Haichuan Tan, Xin Hong, Huaizhi Wang, Tao Huang, Bowen Gao, Yanyan Lan

Organizations: AIPher Technology · Institute for AI Industry Research (AIR), Tsinghua University · School of Pharmaceutical Sciences, Peking University · Department of Computer Science and Technology, Tsinghua University

Abstract

Reverse target fishing identifies protein targets for a query molecule and is central to drug discovery and repurposing. Discriminative approaches adapted from virtual screening face sparse, long-tailed associations and missing-not-at-random measurements: treating unrecorded pairs as negatives can suppress genuine targets. We reformulate the task as generative retrieval and introduce GenFish. Residual vector quantization converts frozen protein language model features into shared semantic identifiers. Given a molecular fingerprint, GenFish autoregressively models codes conditioned on the query and preceding code, ranking proteins by identifier likelihood without explicitly labeling unrecorded pairs as negatives. We introduce ProTargetBench, the first proteome-scale benchmark for reverse target fishing to jointly support measured-panel discrimination and full-catalogue recovery of experimentally recorded positives. Covering 20,431 UniProt-reviewed human proteins, it integrates curated bioactivity evidence and distinguishes experimental inactivity from unmeasured associations. On 198 held-out molecules, GenFish achieves higher mean performance across all three full-catalogue retrieval metrics than four discriminative neural baselines and three non-neural baselines, including ligand-based methods. A controlled positive-hiding experiment shows that targeted removal of false-negative competition improves contrastive retrieval beyond matched random masking; GenFish remains stronger under the same incomplete supervision. Pfam alignment indicates that the semantic identifiers capture biologically meaningful relationships among proteins. These results establish generative retrieval as a principled paradigm for reverse target fishing under incomplete bioactivity records.

Figures & tables

Figure 1: Why sparse reverse target-fishing supervision is difficult. Unmeasured molecule–protein pairs may include both hidden binders and non-binders; treating candidate-level unknowns as negative competitors can suppress undiscovered targets.

Figure 2: Overview of GenFish. (a) Protein tokenization: seven-component protein identifiers are constructed from quantized frozen ESMC-600M features. (b) Molecular prediction: observed positive associations train a molecular predictor to assign likelihoods to target identifiers from Morgan fingerprints. (c) Catalogue retrieval: candidate proteins are ranked by identifier likelihood, with each subsequent semantic code conditioned on the query molecule and the preceding code. Code values are illustrative and do not imply numerical similarity.

Figure 3: ProTargetBench construction and evaluation. The benchmark consolidates bioactivity evidence, fixes a human protein catalogue, and separates full-catalogue positive recovery from measured-panel target ranking.

198 queries $\times$ 20,431 proteins
Method	R@10 $\uparrow$	R@50 $\uparrow$	MRR $\uparrow$
Target frequency	$0.0151$	$0.1364$	$0.0352$
Weighted k-NN	$0.3483$	$0.4187$	$0.4227$
SEA (Morgan)	$0.3074$	$0.4092$	$0.3534$
DrugCLIP	$0.2959\pm 0.0113$	$0.3719\pm 0.0356$	$0.4144\pm 0.0131$
HypSeek	$0.2598\pm 0.0176$	$0.3402\pm 0.0397$	$0.3506\pm 0.0194$

Table 1: Full-catalogue retrieval. All methods rank the same 20,431-protein catalogue for 198 held-out molecules. Neural results are means

\pm

sample SD over five seeds; non-neural baselines report single results. Higher is better (

\uparrow

Measured-panel discrimination
150 binding and 48 enzyme-activity molecules
Method	Binding AP $\uparrow$	Enzyme AP $\uparrow$	Overall AP $\uparrow$
Target frequency	$0.3433$	$0.7212$	$0.4349$
Weighted k-NN	$0.4339$	$0.7152$	$0.5021$
SEA (Morgan)	$0.5020$	$0.7686$	$0.5666$
DrugCLIP	$0.4707\pm 0.0167$	$0.7844\pm 0.0209$	$0.5468\pm 0.0097$

Table 2: Measured-panel discrimination. All methods reuse their catalogue-retrieval scores on the experimentally labeled panels. Overall AP is the molecule-averaged AP across binding and enzyme-activity queries. Higher is better (

\uparrow

Figure 4: False-negative competition ablation. Hidden-positive recovery (left) and held-out retrieval (right) under standard contrastive training, oracle masking, matched random masking, and GenFish with the same visible positives. Bars show means over three seeds with sample-standard-deviation error bars.

Figure 5: Biological structure of the semantic identifiers. Left: Pfam composition of large level-1 code groups. Right: modal-code fractions for prevalent Pfams across semantic levels. Pfam annotations are used only for this post hoc analysis and are excluded from tokenizer and retrieval training.

Appendix figures & tables3 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 6: Pfam correspondence across code levels and catalogue annotation coverage. (a) Fraction of annotated protein pairs sharing a code that also share at least one Pfam. Thin lines show five tokenizer seeds; markers and error bars indicate their mean and sample standard deviation. The vertical axis is logarithmic. The dashed line shows the fraction among all annotated catalogue pairs. The suffix resolves identifier collisions and is not a semantic level. (b) Catalogue protein counts by number of distinct Pfam annotations; zero indicates missing cross-references.

Method	Learning rate	Dropout	Temperature	Parameters	Optimizer
DrugCLIP (adapted)	$5\times 10^{-5}$	0.1	0.03	2,164,736	AdamW
HypSeek (adapted)	$10^{-4}$	0.1	0.03	2,164,736	AdamW
LigUnity (adapted)	$5\times 10^{-5}$	0.1	0.07	4,525,570	AdamW
ConPLex (no-DUDE)	$10^{-4}$	0.0	–	3,147,776	AdamW
GenFish	$10^{-4}$	0.0	–	3,887,864	Adam

Appendix

Table 3: Validation-selected headline configurations. Temperature applies only to contrastive objectives. Parameter counts refer to the trainable predictors and exclude frozen protein feature extractors.

Retrieval by training-ligand similarity
Maximum Morgan-fingerprint Tanimoto similarity; median 0.511
Method	$[0,0.4)$ , $n=43$	$[0.4,0.6)$ , $n=79$	$[0.6,1]$ , $n=76$
R@10 $\uparrow$
Target frequency	$0.0000$	$0.0248$	$0.0135$
Weighted k-NN	$0.0777$	$\mathbf{0.4332}$	$0.4130$
SEA (Morgan)	$0.0667$	$0.4059$	$0.3411$

Appendix

Table 4: Full-catalogue retrieval stratified by maximum training-ligand Tanimoto similarity. All methods reuse the scores in Table 1 , without refitting. Neural entries report five-seed means

\pm

sample SD; other entries are single results. Higher is better for all metrics (

\uparrow

); bold marks the highest mean within each metric and similarity group.

Drug Discovery

Drug discovery identifies, designs, and tests molecules that can safely affect disease biology.

See all 164 papers

Jun 4, 2026

ShallowBench: Benchmarking Generative Drug Design Models on Shallow-Pocket Targets

While generative AI models have demonstrated remarkable success in structure-based drug design, they predominantly rely on deep binding pockets and struggle to sample effective ligands for challenging low-pocketability targets, such as the historically "undruggable" oncology targets KRAS and MYC. To address this gap, we introduce ShallowBench, a strictly curated benchmark of 5,780 shallow-pocket targets extracted from CrossDocked2020. By computing the difference between an Alpha Shape "lid" volume and the underlying protein atom voxel volume, we successfully isolated targets with low concavity while ensuring sufficient surface area for binding. Evaluating various state-of-the-art generative models reveals weaker predicted binding affinity on these low-concavity interfaces. ShallowBench therefore provides a rigorous benchmark for generative biology models and highlights the necessity of new architectural innovations or loss functions capable of navigating these challenging targets.

Saket Reddy, Shiwei Liu

University of Illinois - Urbana-Champaign

Benchmark Design Structure-Based Drug Design Generative Modeling Molecular Generation Drug Discovery

Ask PDF

May 11, 2026

Beyond Manual Curation: Augmenting Targeted Protein Degradation Databases via Agentic Literature Extraction Workflows

Predictive models in biomedicine depend on structured assay data locked in the text, tables, and supplements of primary publications. This bottleneck is especially acute in targeted protein degradation (TPD), where each assay record must combine compound identity, degradation target, recruiter, assay context, and endpoint values reported across sections, tables, and supplementary files. Inconsistent compound identifiers and incomplete or implicit assay context further demand domain-specific logic that generic LLM pipelines do not provide. Existing molecular glue and PROTAC databases are manually curated and often lack the experimental context required for downstream modeling. We formulate TPD database extraction as a domain-specific curation task and present an expert-in-the-loop LLM workflow, evaluated through a triangular comparison among LLM predictions, standardized baseline records, and expert-annotated ground truth. A lightweight cross-validated prompt-refinement module adapts extraction instructions from scarce expert annotations. With only seven annotated molecular glue publications, the workflow achieved record-level

F_1 = 0.98

and transferred to PROTACs by terminology substitution alone, maintaining record-level

F_1 > 0.93

. Applied at scale, it expanded molecular glue and PROTAC databases by 81% and 92% records, respectively, with 92% and 82.5% of newly recovered records validated as correct upon expert review. The workflow also recovered kinetic and assay-context information essential for cross-study potency comparison and condition-aware degradation modeling. We release the workflow, prompts, evaluation code, and extracted datasets as resources for TPD data curation and AI-assisted scientific curation more broadly.

Yaochen Rao, Farzaneh Jalalypour, N. M. Anoop Krishnan +1

Department of Computer Science and Engineering, Chalmers University of Technology and University of Gothenburg, Gothenburg, SE · Yardi School of Artificial Intelligence, Indian Institute of Technology Delhi, New Delhi, IN

Training Data Curation Document Information Extraction LLM Information Extraction

Ask PDF

Oct 14, 2026

InstructPro2: Natural Language and Biological Grounding Guided De Novo Protein Design

The ability to generate proteins with desired functions is a central goal of computational biology. Although natural language can express diverse design objectives, existing protein generation methods often support only narrow tasks or make limited use of biological constraints. We develop InstructPro2 to translate natural-language design objectives into protein sequences while incorporating task-specific biological context. A text encoder captures the design instructions, while dedicated ligand and antigen encoders provide biological context to the sequence-generating decoder. For pretraining, we assemble InstructPro2Bench, comprising 205.7 million <function description, protein> pairs from UniProtKB. We subsequently adapt the model to ligand-binding protein and antibody design. On functional protein generation benchmarks, InstructPro2 outperforms Pinal, the strongest baseline, achieving a functional recovery F1 of 74.8 versus 62.2 and lowering global RMSD from 5.33A to 3.04A. For antibody design, the model obtains a DockQ score of 0.39, while ligand-binding protein design reaches a docking success rate of 72.6%. These findings highlight how integrating language-based instructions with biological conditioning can extend protein generation to distinct design settings.

Charles Novak, Zhenqiao Song, Chuan Li +2

Ray and Stephanie Lane Computational Biology Department, School of Computer Science, Carnegie Mellon University, Pittsburgh, PA 15213, USA · Language Technologies Institute, School of Computer Science, Carnegie Mellon University, Pittsburgh, PA 15213, USA · Lambda Inc, San Francisco, CA 94105, USA

Protein Design Protein Language Models Protein Binder Design Conditional Generative Modeling

Ask PDF

Abstract

Drug Discovery

Explore similar work

ShallowBench: Benchmarking Generative Drug Design Models on Shallow-Pocket Targets

Beyond Manual Curation: Augmenting Targeted Protein Degradation Databases via Agentic Literature Extraction Workflows

InstructPro2: Natural Language and Biological Grounding Guided De Novo Protein Design