Texas A&M University researchers have released an artificial intelligence model that promises to cut one of the most frustrating drains on tuberculosis drug development: false-positive screening hits that look promising but waste months of laboratory work. Dubbed CAGE-Fusion, the model distinguishes between genuinely active compounds and four types of “nuisance” molecules that commonly trick early-stage assays. In head-to-head tests, it correctly flags the nuisance candidate about 94% of the time, according to a study published this week in the Journal of Cheminformatics. The tool does not discover new drugs itself; instead, it acts as a triage layer, helping chemists decide which compounds are worth the most expensive follow-up experiments.

Inside CAGE-Fusion: A Binary Classifier for Nuisance Compounds

CAGE-Fusion—short for Co-Attention Graph Embedding Fusion—is a deep learning model that evaluates candidate small molecules using two complementary data representations. It ingests both the molecule’s graph structure (atoms as nodes, bonds as edges) and its SMILES string (a machine-readable chemical notation). A gated co-attention mechanism allows the model to weigh the two views against each other, surfacing the most suspicious features regardless of which representation they come from.

The model targets four well-known categories of assay interference:
- Aggregators: Compounds that clump into colloidal particles, nonspecifically inhibiting protein activity.
- Fluorescence interferers: Molecules that mimic or quench the fluorescent signal used in many screening assays.
- Redox cyclers: Reactive compounds that generate false signals by cycling between oxidation states.
- Promiscuous binders: Molecules that stick to many protein targets rather than the one under study.

Importantly, CAGE-Fusion does not just return a risk score. The team built in an explainability layer that highlights specific molecular substructures driving the classification. “The model can walk you through the process and show you which regions of the molecules it thought were problematic,” said Siddhant Rath, the AgriLife Research scientist who led the work. For an organic chemist staring at a structure diagram, that visual feedback is more valuable than a black-box probability—it suggests where to tweak the molecule to fix the problem.

Performance metrics from the peer-reviewed paper place the model’s macro-averaged ROC-AUC at 0.94 on the nuisance-compound classification task. That number measures how well the model separates the interference classes across all categories, not how often a flagged compound will fail in real follow-up tests. Still, in practical ranking exercises, given one clean compound and one nuisance, CAGE-Fusion picked the nuisance as more suspicious 94% of the time. The model is especially adept at catching redox cyclers and somewhat weaker at flagging promiscuous binders, the researchers note—a pattern that mirrors the relative noisiness of those classes in available training data.

The Platform Behind the Model: DAIKON

CAGE-Fusion does not operate in isolation. It plugs into DAIKON, an open-source data-management platform that Sacchettini’s lab first released in 2023 and that is now used by the Tuberculosis Drug Accelerator (TBDA), a consortium of labs and pharmaceutical companies backed by the Bill & Melinda Gates Foundation.

DAIKON tracks a drug target from its genetic underpinnings through years of medicinal chemistry. Every synthesis, assay result, and decision point gets recorded in one searchable system. That structure solves a knowledge-management headache familiar to anyone who has ever hunted through network drives, slide decks, and email threads for a vital piece of data. “In my research, I might come across a molecule and think, ‘I know I’ve seen this before,’ but there was no easy way to find it,” Sacchettini said. DAIKON changes that.

The platform’s newest additions include a chat-style interface that lets researchers query TBDA’s accumulated records in natural language. “It will tell me in a matter of seconds who presented something, what they said about it, and give me the presentation so I can see the slides,” Sacchettini added. For enterprise IT professionals, this will sound like a domain-specific SharePoint or Confluence instance supercharged with retrieval‑augmented generation—turning institutional memory into something actually retrievable.

All of Sacchettini’s recent AI tools, including CAGE-Fusion and the chatbot-like query system, run inside DAIKON. That design choice is deliberate: rather than forcing scientists to ping a separate web service with SMILES strings, the tools sit next to the data they analyze. When a new screening campaign uploads results, CAGE-Fusion automatically inspects the hits and flags likely nuisance compounds before a human ever decides to pour them into the next assay plate.

Why TB Drug Discovery Wastes So Much Time on Dead Ends

To appreciate what CAGE-Fusion and DAIKON actually accomplish, it helps to understand why tuberculosis drug discovery is so painfully slow. Mycobacterium tuberculosis, the bacterium that causes TB, possesses a thick, waxy cell envelope that blocks most small molecules from entering. The organism also grows extraordinarily slowly—doubling times of 24 hours versus 20 minutes for common lab bacteria like E. coli. “To do a tuberculosis experiment takes months sometimes, where it can take a week with staph or strep,” Sacchettini explained. That tempo means every false lead costs not just money but countable weeks of limited BSL‑3 facility time.

The problem of nuisance compounds is not unique to TB, but the cost of pursuing them is magnified. An early screen against a protein target might spit out 10,000 hits. After computational filters, a few hundred might look promising, and a dozen might be selected for time‑consuming validation assays. If half of those are assay artifacts, the team has just burned half a year’s capacity on dead ends. CAGE-Fusion’s job is to push those artifacts to the bottom of the priority list before the validation phase begins.

Who Really Benefits—and How

For TB researchers The immediate practical use is clear. Anyone participating in the TBDA can access CAGE-Fusion through DAIKON and apply it to their own screening data. The output is a reranked list of hits, with the most likely interferers demoted and annotated with visual explanations. The model is not yet a standalone command-line tool that outsiders can download and run on arbitrary datasets, but its underlying architecture is described in enough detail that a competent cheminformatics group could reproduce it.

For IT and data professionals in life sciences The Texas A&M story is a case study in architecting AI that sticks. Rather than building a flashy model and hunting for a problem, the lab first built (and socialized) the data infrastructure that made the model useful. DAIKON handles versioning, access control, and structured metadata—exactly the boring plumbing that determines whether an AI initiative delivers ROI or gathers digital dust. The explainability component of CAGE-Fusion also addresses a major barrier to adoption in regulated environments: if a model merely spits out “98% probable artifact,” a skeptical chemistry team will ignore it. By pointing to specific problem moieties, the tool gives scientists a reason to trust—or challenge—the verdict.

For the broader public Any technology that speeds up TB drug discovery matters because TB remains the world’s deadliest infectious disease, killing roughly 1.5 million people each year. Multi‑drug‑resistant strains complicate treatment, and the current first‑line regimen takes six months of daily antibiotics. Shorter, more effective therapies depend on finding new chemical starting points—a task made easier when screening hits are more trustworthy. CAGE-Fusion will not deliver a new drug by itself, but every month of lab work it saves moves a potential new drug one month closer.

How We Got Here: A Timeline

The Sacchettini lab has worked on TB drug discovery for decades, including solving the structure of key M. tuberculosis targets. The pivot toward data infrastructure came with:
- 2023: DAIKON open‑source release, designed to consolidate fragmented research records across the TBDA consortium.
- 2024–2025: TBDA members populate DAIKON with years of historical chemistry and biology results. The lab begins experimenting with machine learning models trained on this unified dataset.
- Early 2026: Development of CAGE‑Fusion, leveraging graph neural networks and SMILES‑based encoders. The model trains on publicly available assay‑interference labels plus in‑house data.
- July 2026: Peer‑reviewed publication in the Journal of Cheminformatics, accompanied by announcements in News‑Medical and Drug Target Review.

The funding came mostly from the Gates Foundation, with additional support from the Welch Foundation and the National Institutes of Health. That philanthropic backing is key: TB drug research is not a blockbuster commercial market, so academic and nonprofit labs carry much of the load.

What to Do Now

If you work in a drug‑discovery consortium or an academic screening center, here are concrete steps:

  1. Evaluate your hit‑triage pipeline. If you currently rely on manual substructure filters or simple rule‑based flagging (like PAINS filters), compare their performance against CAGE‑Fusion’s reported numbers. The model’s ability to catch aggregators and redox cyclers may outperform legacy filters, especially for unusual chemotypes.
  2. Consider your data fabric. DAIKON’s success suggests that AI models are only as good as the structured data they sit on. If your lab’s knowledge lives in PowerPoint and Excel, invest in a centralized, queryable record system before you buy into fancy ML.
  3. Demand explainability. When evaluating third‑party AI tools, ask for visualization of why a compound was flagged. Black‑box rejections will erode trust fast. CAGE‑Fusion’s attention‑based highlighting is one pattern worth copying.
  4. Explore open‑source alternatives. The DAIKON platform is open‑source. Even if you don’t work on TB, its data model may be adaptable to other neglected‑disease programs.
  5. For the curious tech builder: The model architecture is described in the paper. With sufficient computational chemistry skills, you could retrain a similar fusion model on your own nuisance‑compound dataset.

What’s Next

CAGE‑Fusion is entering the real‑world proving ground inside TBDA. Over the next year, consortium members will see whether the model’s flags consistently reduce wasted validation assays. If it performs well, the approach could be extended to other bacterial or parasitic diseases that share the slow‑growth problem, such as leprosy or Buruli ulcer.

Meanwhile, Sacchettini’s group is not stopping at nuisance‑compound detection. They are working on AI tools that integrate biological pathway knowledge, predict toxicity, and suggest synthetic routes—all inside DAIKON’s collaborative framework. The vision, as Sacchettini puts it, is not “AI will give us the exact right answer” but “it can tell us what not to work on, which then informs us what we should be working on.” That humility, paired with rock‑solid data plumbing, may be the real lesson for anyone trying to put AI to work in a messy, expensive real‑world domain.