AI's Role in Improving Health and Well-Being for Louisiana and Beyond

Chapter three of this AI series covers nine health research programs. Several focus on drug discovery and antibiotic resistance. Others address cancer treatment, neurological disease, and pandemic preparedness. Two were designed specifically for conditions in Louisiana's rural hospitals. All of them began with an unmet clinical need rather than with a technology in search of an application.

DeepDrug: Rebuilding Drug Discovery from the Ground Up
Drug development is expensive and slow. Some estimates place the average capitalized cost of bringing a new drug to approval at more than a decade and several billion dollars, and the majority of candidates fail before reaching patients. DeepDrug was designed to improve the early-stage screening step so that more resources flow to compounds with genuinely stronger properties.

Louisiana communities carry health burdens that are measurable and specific: cancer rates above the national average in several parishes, antibiotic-resistant infections in hospitals, neurological diseases that have few effective treatments, and rural areas where specialists are a long drive away. Some of the drugs these patients need do not yet exist, or have not been developed because the economics favored more profitable conditions.

DeepDrug addresses the front end of that problem. The platform uses AI to evaluate millions of drug candidates computationally before laboratory resources are committed. One module breaks known drugs into their molecular components. A second recombines those components to generate new candidate molecules. A third screens each candidate for toxicity and manufacturability. A fourth evaluates drug combinations that may work more effectively together than individually.

Working through this pipeline, the platform can take a disease target and return a ranked list of candidates in weeks rather than the years that traditional screening programs require. The highest-ranked results are routed to LSU Health researchers for experimental testing with Louisiana patients.

The IBM Watson AI XPRIZE, a competition among 147 international teams working on large-scale real-world AI applications, named our team a semifinalist. The program has also initiated drug repurposing programs targeting ALS and COVID-19, described later in this chapter.

 

MDF-DTA: Predicting How Well a Drug Will Bind
Effective drug action often depends on strong and selective interaction with a biological target, although binding affinity is only one part of overall drug performance. Predicting that binding computationally, before synthesis, is one of the harder problems in early-stage drug development.

Every drug works by interacting with a specific protein in the body. The fit between molecule and protein matters: too weak and the drug has little effect; too nonselective and it hits unintended proteins and causes side effects. Most earlier computational methods for predicting this fit read drug molecules in only one format, which means they capture some aspects of the molecule's behavior while missing others.

Our MDF-DTA system reads three representations simultaneously: the molecule's chemical structure written as a text string, its atom-by-atom bond network, and its three-dimensional shape. Each format reveals something the others do not. Combining them gives the model a richer picture of how a molecule is likely to interact with its target protein.

We published MDF-DTA in the Journal of Chemical Information and Modeling in 2024. On the standard benchmark datasets for drug-target binding prediction evaluated in our study, MDF-DTA achieved stronger performance than the comparison methods reported at the time of publication. It now serves as the core scoring engine inside DeepDrug, producing the binding estimates that drive candidate ranking across the platform.

 

New Antibiotics for Drug-Resistant Infections
MRSA and other drug-resistant pathogens are a documented problem in Louisiana hospitals. Standard antibiotic development has slowed significantly because the economics are difficult for pharmaceutical companies. Our team is pursuing a computational approach to generating new compound candidates that work against strains existing drugs cannot reach.

Patients with drug-resistant infections stay in hospital longer, face higher treatment costs, and die at higher rates than patients with susceptible infections. The pipeline of new antibiotics being developed commercially has narrowed considerably over the past two decades, leaving clinicians with a limited and slowly shrinking set of options.

Our approach disaggregates known antibiotic structures into their molecular components and recombines them in new configurations. Because molecules built this way can be structurally unlike existing antibiotics, they may initially avoid some established resistance mechanisms, although laboratory and clinical studies are needed to assess whether and how quickly resistance could develop.

The AI screening step includes an explanation layer that highlights which parts of each candidate molecule drive its predicted activity. The chemistry team can use that output to strengthen those structural features before committing to expensive synthesis steps.

We presented this program at the Florida AI Research Symposium in 2026. Computationally identified anti-MRSA compound classes are advancing toward preclinical in vitro testing, which will determine whether the computational predictions hold under laboratory conditions.

 

Trans-ARG: Identifying Resistance Genes Before Choosing a Treatment
The question this system addresses is straightforward: which antibiotic resistance genes does this bacteria carry? The answer determines which drugs are likely to work. Trans-ARG was built to produce that answer more reliably, drawing on the largest unified resistance gene dataset assembled for this purpose.

When a patient presents with a serious bacterial infection, treatment selection depends heavily on knowing which resistance genes the pathogen carries. The wrong antibiotic not only fails to clear the infection but applies selection pressure that can accelerate resistance development. Rapid and accurate resistance gene identification is particularly important in rural Louisiana hospitals, which often lack the specialized microbiology staffing available at academic medical centers.

We built Trans-ARG by first assembling a unified collection of nearly 39,000 labeled resistance gene sequences drawn from nine scientific databases that had been maintained separately, creating a training set substantially larger than those used by earlier tools. We then trained the model using a transfer learning approach: the model first learned general protein sequence patterns from 250 million protein sequences across the tree of life, then applied that foundational knowledge to the specific task of resistance gene classification.

On the standard benchmark that distinguishes true resistance genes from structurally similar but harmless sequences, Trans-ARG achieved 97 percent precision-recall F1 on the held-out evaluation set. Overall resistance gene classification accuracy across the dataset was above 90 percent. In the comparison included in our 2024 publication, Trans-ARG produced stronger benchmark results than the reference tools evaluated alongside it, including systems from international genomic databases.

Improving resistance-gene identification is relevant for infection-control teams in any hospital but is especially valuable in settings where specialist consultation is not immediately available and where the speed of classification directly affects treatment decisions.

 

Head and Neck Cancer: Getting the Treatment Right the First Time
Two patients with what looks like the same tumor can respond completely differently to the same treatment. The question our team is working on is whether the biological profile of a tumor, read before treatment starts, can predict which approach a specific patient's cancer is most likely to respond to.

Head and neck cancer covers tumors of the mouth, throat, and larynx. Louisiana patients develop these cancers at rates above the national average. The treatments available, surgery, radiation, and chemotherapy, are effective for some patients but carry lasting consequences: permanent swallowing difficulty, dry mouth, and in some cases disfigurement. Getting the right treatment the first time matters both clinically and in terms of quality of life.

Our team, working in close partnership with LSU Health oncologists, is developing AI that reads the full biological profile of a tumor before treatment begins. The model draws on genetic data from the tumor's DNA, gene expression patterns showing which genes are actively being used, epigenetic markers that modify gene activity, and the patient's clinical history. It trains on records from past patients with known outcomes and then applies those patterns to predict which treatment a new patient's specific tumor is most likely to respond to.

We have also applied AI to characterizing the immune environment surrounding tumors. The composition of that environment influences whether a patient is likely to respond to immunotherapy. This analysis is helping to identify markers that could guide patient selection for clinical trials being developed with LSU Health.

These models are still in development and have not yet been validated in prospective clinical settings. The goal is to generate enough evidence to support clinical trial design in partnership with LSU Health oncology teams.

 

Answer ALS: Taking on a Disease Medicine Has Not Solved
ALS has two FDA-approved treatments, neither of which does much. Most patients die within two to five years of diagnosis. Our program runs four parallel AI approaches on multiple large datasets to identify biological targets and drug candidates, with clinical trials at LSU Health as the intended endpoint.

Part of why ALS has been so resistant to treatment is that it is not one disease. Different patients have different underlying biological drivers. A treatment that helps one subgroup may have no effect on another. Standard clinical trials, which tend to enroll ALS patients without stratifying by biological subtype, may miss effects that are real but limited to specific populations.

Our program draws on two major data sources. The Answer ALS consortium has assembled patient-derived biological samples and clinical data from a large multi-institution collaboration. The ALS Therapy Development Institute database adds further depth. We apply DANN modeling to process genetic, gene expression, protein, and metabolic data simultaneously, identifying which biological pathways are most disrupted and which proteins are the most tractable targets for intervention. Analysis of GATC Health data has produced an initial set of high-priority targets now moving into the testing queue.

With targets identified, we run two parallel strategies. The first models T cell immune responses using those target proteins as inputs, looking for signals that might direct immune activity against the disease process. The second pursues drug repurposing through four AI methods running simultaneously: matching existing FDA-approved drugs to the three-dimensional structure of ALS target proteins, estimating binding strength, generating structured explanations of why specific compounds look promising, and evaluating combinations that might work synergistically.

The goal is to generate sufficient evidence for NIH funding and pharmaceutical partnerships that would bring early clinical trials to LSU Health. Running these approaches in parallel, across multiple datasets, compresses a timeline that would otherwise extend over many years of sequential investigation.

 

eVir: Preparing Louisiana for the Next Pandemic
When SARS-CoV-2 emerged, there was no fast way to identify which existing approved drugs might be effective against it. We built a tool called eVir to address that gap. Applied to COVID-19, it identified two candidates, one of which reduced viral load in animal studies with no observed toxicity.

Louisiana was hit hard by COVID-19, particularly in rural communities with fewer healthcare workers and limited access to hospital-level care. The underlying challenge, however, is not unique to COVID-19. Every time a new pathogen emerges, the search for effective treatments begins largely from scratch. Developing a genuinely new drug takes a decade or more. But thousands of safe, approved drugs already exist, some of which almost certainly have activity against any given new pathogen. The bottleneck is identifying which ones, quickly enough to be useful.

eVir maps the interactions between drugs and the human proteins they bind to, then uses AI to find drugs whose interaction patterns resemble those of natural antiviral proteins the immune system deploys against viruses. The idea is that a drug already known to engage the same proteins an antiviral protein engages may have antiviral properties worth testing.

Applied to SARS-CoV-2, eVir identified two candidates. The first was Imatinib, a leukemia drug, flagged for its predicted ability to dampen the inflammatory overreaction responsible for severe COVID-19 outcomes. The second was Mebendazole, an antiparasitic medication widely available and inexpensive in the developing world, flagged for predicted direct antiviral activity.

Animal studies supported the Mebendazole prediction: mice treated with the drug showed significantly lower viral loads in lung tissue, with no observed toxicity or organ damage. These are encouraging results, though animal study findings do not always transfer to human patients and further clinical evaluation would be needed before any treatment conclusion could be drawn.

Because Mebendazole is already approved, affordable, and globally available, it may be particularly relevant for rural Louisiana communities and for lower-income countries that cannot access newly developed antivirals at scale. eVir can be directed at other emerging pathogens as needed.

 

Catching Dangerous Drug Reactions Before They Harm Patients
Harmful unintended drug reactions injure and kill hospital patients at rates that have been difficult to reduce through manual review alone. The information needed to catch many of these events is sitting in clinical notes written every shift. Our system reads those notes continuously and provides an evidence-based explanation for each alert it generates.

Estimates from the published literature place the annual number of serious adverse drug events in U.S. hospitals at over two million, with more than ten thousand deaths attributed to drug reactions each year. Louisiana's situation is particularly difficult in rural hospitals, where clinical pharmacists are often unavailable overnight and during weekends. The free-text notes that physicians and nurses write each shift contain the information needed to identify dangerous drug-reaction patterns, but extracting it reliably had been beyond the capacity of earlier automated systems.

Our pharmacovigilance team, working with Oracle India and IIIT-Delhi, built a system that reads clinical notes in real time and flags potentially dangerous drug combinations as they appear in the record. The system pairs a large language model with a structured knowledge base of documented drug interactions. When the model identifies a drug mentioned alongside a potential adverse effect, it checks against the knowledge base, traces the supporting evidence, and generates an alert with a plain-language explanation that includes the drug, the suspected reaction, and the evidence chain.

This design reflects a clinical constraint: a clinician is unlikely to act on an alert that arrives without an explanation they can evaluate. Providing supporting context with each alert allows the receiving clinician to assess the recommendation rather than having to accept it on the basis of an unexplained score.

In our evaluation, the system correctly identified dangerous drug-reaction pairs in more than 94 percent of cases on the test set. Pairing the language model with structured biomedical knowledge improved performance by 8 to 16 percent relative to either approach used alone, suggesting that neither component is sufficient on its own. These results were obtained on held-out test data; real-world deployment performance would need to be validated separately at each clinical site.

The system was specifically designed for hospitals with limited overnight pharmacy staffing. That constraint shaped the design from the start rather than being addressed as an afterthought.

 

Protecting Infants Who Cannot Cry for Help
Infants with tracheostomy tubes cannot vocalize when something goes wrong. Standard monitoring detects oxygen loss after the airway event has already begun. We are developing a bedside video monitoring system designed to detect signs of tube displacement or distress earlier and alert clinical staff. The system was built with rural Louisiana hospitals as the primary deployment context.

A tracheostomy tube is a small surgical airway placed through the neck. Some infants require one to survive. Because it bypasses the vocal cords, a baby with a tracheostomy cannot make a sound when the tube becomes dislodged or blocked. Standard pulse-oximetry monitoring catches the resulting oxygen drop, but by the time that signal appears, the response window is already narrowing. This risk falls hardest on infants in smaller and rural hospitals, where continuous nursing observation is not consistently available.

We are developing an AI-powered bedside video monitoring system intended to close that gap. The system uses a small camera at the infant's crib and runs two models simultaneously. The first watches the infant's neck and classifies what it sees into four clinically relevant categories related to tube position. It produces a classification in under five milliseconds. The second model watches the infant's face and distinguishes signs of distress from calm states, achieving close to 90 percent classification accuracy on the validation set under controlled test conditions. Both outputs are combined with the infant's oxygen readings to create a layered alert signal.

The system is calibrated during the first ten days of deployment using observations from the specific infant. Each baby's anatomy differs in terms of skin tone, tube size, and behavior patterns. A generic model without calibration would generate too many false alerts, which clinical staff would quickly learn to disregard. A calibrated model generates alerts with a higher likelihood of being actionable.

All video processing runs on an encrypted device at the crib. No video data leaves the room. This privacy-first architecture was chosen specifically because rural Louisiana hospitals often lack the data governance infrastructure available at larger institutions, and any system requiring external data transmission would face significant barriers to deployment there.

The system is still in development. Testing in clinical settings has begun, but validation in a range of rural hospital environments, across patient populations, and over extended deployment periods will be needed before the system could be considered ready for broad adoption.

 

What These Nine Projects Have in Common
All nine projects are grounded in a specific problem that is affecting real patients in Louisiana. All involve teams that span multiple disciplines. None of them started with an AI capability and then looked for a problem to apply it to. They started with the problem and built the technology around what the problem required.

Drug discovery. Antibiotic resistance. Cancer treatment. ALS. Pandemic preparedness. Drug safety. Infant monitoring. These are not abstract research areas. They are specific gaps where Louisiana patients are underserved and where improved tools would have a measurable effect on outcomes.

The ninth project in this list, the infant monitoring system, reflects something I think generalizes across the program. It was designed for a rural Louisiana hospital with limited staffing and limited data infrastructure. Designing for those constraints forced clarity about what the system actually needed to do and what assumptions it could not safely make. A system that performs reliably in that setting may also perform well in better-resourced environments, though that would need to be confirmed through evaluation at each additional site. That is what designing for the hardest case looks like in practice.