Psst… We received a tip…

Somewhere, a patient is waiting for a drug that does not exist yet.

Finding useful medicine is brutally hard, and biology has a nasty habit of hiding important clues.

Picture a researcher trying to design a new drug while buried under papers, protein databases, lab results, and old trial data. The answer might already be out there, but you still need the right system to pull it in at the right moment. 

Generative AI can already help propose new proteins and drug-like molecules. But drug design is an iterative loop of designing, checking, comparing, revising, and repeating. The challenge is pulling in the right context at the right moment, so the model can update weak assumptions before they become expensive dead ends.

Deep EigenMatics’ patent points at that exact problem. An AI drug-design system that can update what it knows while it is working, giving tomorrow’s medicines a better shot at being built from the right context.

This week’s patent was filed by Deep EigenMatics LLC, an AI-biotech company based in Texas. The patent is called ‘CLUE: Dynamic Context Retrieval in Reasoning Models for AI-Based Protein and Drug Design’.

The name sounds like a lab acronym because it is one. CLUE stands for Context Load Update Engine. This is an AI system that helps design the building blocks of possible new medicines, while checking for better information as it works. 

Instead of just taking a prompt, searching once, and then spitting out a final answer, it starts with a request, pulls in relevant outside information, begins generating a design, checks whether it needs better context, searches again if needed, and then keeps going.

Let’s work through an example of how a scientist could use this tech in Alzheimer's drug design. 

… But the model does not have to finish the whole thing in one pass. 

That is what makes CLUE different from a normal retrieval system. A normal retrieval system usually searches at the beginning. CLUE can search again during generation, based on what the model has already started to produce.

AlphaFold cracked one of biology's hardest puzzles. Given a protein's genetic sequence, it predicts the exact shape it folds into. That had stumped scientists for fifty years. Google DeepMind, the London-based AI research lab behind AlphaFold, essentially solved it in 2020. But predicting how a protein folds is not the same as designing one from scratch to do a specific job. That problem is still wide open, and a lot of money is chasing it.

A 2021 BIO report found that less than 1 out of 10 of the drug candidates entering Phase I clinical trials eventually reached approval, and Deloitte put the average 2024 R&D cost per asset at about US$2.23 billion. (BIO) (Deloitte)

Drug design has too many expensive dead ends. The weak point is speed and waste. A bad early guess can burn years of lab work, investor money, patient hope, and researcher attention before everyone admits the molecule was never good enough.

The perfect one-for-one competitor is hard to find, because this patent is about a specific retrieval loop inside the model. But the broader race of trying to make drug discovery less wasteful with AI is packed.

Schrödinger leans on physics-based modelling and software that drug companies already pay to use. Schrödinger spent about US$201.8 million on R&D in 2024 and says its molecular discovery platform is licensed by biotech, pharma, industrial, and academic customers. Isomorphic Labs, the Google DeepMind spinout, raised US$600 million in 2025 to push its AI drug-design engine forward. (Schrödinger) (Isomorphic Labs)

Most players are attacking the problem by building bigger platforms. Recursion maps biology with huge cell-imaging datasets. (Recursion)

The closer technical cousins are coming from research labs. Rag2Mol uses retrieval-augmented generation (RAG) to design small molecules for a 3D protein pocket. It is not the same as CLUE, but it shows the same direction of travel, that models designed from memory alone are old news. (Briefings in Bioinformatics)

The FDA needs to brace for the floodgates to open. 

It is predicted that Phase I success rates for AI-designed drugs may reach 80–90%, compared to 40–65% for traditionally discovered drugs, with overall timelines running 40% faster and preclinical costs reduced by 30–70%. If this holds at scale, the number of drug candidates credibly reaching Phase I, and subsequently Phase II, could surge well beyond anything the regulatory system was designed to absorb. (Axis Intelligence)

The institution on the other side of that flood is the FDA. The FDA's drug review division, the Center for Drug Evaluation and Research (CDER), employed approximately 7,580 staff prior to recent restructuring. That workforce has since turned sharply. CDER lost 746 staff in the final quarter of fiscal year 2025 alone, gaining just six, erasing all net staffing gains from 2023 and 2024. 

Meanwhile, a related executive order limits future federal hiring to one new employee for every four who depart. The FDA is, in short, moving in the opposite direction to the industry it regulates. (AgencyIQ by POLITICO)

The Department of Government Efficiency-directed cuts have already caused drug reviews to miss decision deadlines. The bottleneck will be at the regulatory gate, staffed by an institution that is still trying to define the rules of the game while the floodwaters rise.

Bringing AI to an AI fight. 

One proposed answer to the floodgates problem is to meet AI with AI. 

The European Medicines Agency released a 2025–2028 workplan to leverage large volumes of regulatory and health data, essentially deploying AI to review AI-assisted submissions. 

In May 2025, the FDA Commissioner announced the use of AI in the submission review process, following a generative AI pilot program for scientific reviewers that the agency described as successful. 

But the feasibility of this approach runs into an uncomfortable paradox. Early implementation of AI within regulatory bodies has already revealed key challenges around accuracy, consistency, hallucinations and misrepresentations. This is a serious problem for review institutions, and deeply troubling when the output is a safety determination on a novel therapeutic. (White & Case)

We already know this is one of the biggest markets in the world…

The market behind this patent is enormous because medicine itself is enormous. Evaluate forecasts global prescription drug sales to top US$1.7 trillion by 2030. (Evaluate)

IQVIA expects annual medicine use to reach 3.8 trillion defined daily doses by 2028, while global medicine spending at list prices is forecast to rise 38% through 2028.

The R&D machine is already huge. EFPIA’s 2025 industry report says health industries invested about €258.1 billion in R&D in 2023, making up 20.5% of total business R&D spending worldwide.

Now zoom into AI drug discovery. Grand View Research estimates the global AI-in-drug-discovery market at US$2.35 billion in 2025, growing to US$13.77 billion by 2033. Global Market Insights is even more bullish, putting the market at US$3.1 billion in 2025 and forecasting US$43.9 billion by 2035.

McKinsey has estimated that generative AI could unlock US$60 billion to US$110 billion a year in value for pharma and medical products. And drug discovery is one of the places where better models could actually move the needle.

PitchBook says global VC deal value in AI drug discovery reached US$3.8 billion in 2025, the second-highest annual level on record.

Xaira Therapeutics launched with more than US$1 billion in funding, backed by names like ARCH Venture Partners, Foresite Capital, Sequoia, NEA, Lux, Lightspeed, Menlo, and others. Xaira Therapeutics is building an AI-driven drug discovery company that combines machine learning, large-scale biological data generation, and drug development to create new medicines.

Pathos AI raised US$365 million in 2025 at a roughly US$1.6 billion post-money valuation to build an AI platform for oncology drug development, using clinical, molecular, and imaging data.

Chai Discovery is a smaller but cleaner signal. It raised US$70 million in Series A funding, bringing total funding to US$100 million, after showing progress in AI-designed antibodies. Its Chai-2 model reached a near 20% hit rate in de novo antibody design, compared with 0.1% for the previous computational state of the art. (Business Wire)

The deal flow is also shifting from pure software into automated drug-making systems. Excelsior Sciences raised US$95 million in 2025 to use AI and machines to speed small-molecule development, with Reuters reporting that investors believe a typical four-month process could shrink to about two weeks in a single automated facility. (Reuters)

In big pharma, Eli Lilly expanded its partnership with Insilico Medicine in a deal worth up to US$2.75 billion, including a US$115 million upfront payment, to use Insilico’s AI engine and license certain preclinical oral drug candidates. (Reuters)

Recursion agreed to acquire Exscientia in a US$688 million all-stock deal, combining Recursion’s biology platform with Exscientia’s chemistry design and small-molecule capabilities. When startups start merging like this, it usually means the market is leaving the toy phase and entering the “who has the full stack?” phase. (Reuters)

The hopeful version of this story is that an AI drug-design system knows when to stop, check its notes, pull in better context, and keep going.

That matters because the future of medicine may come from systems that can stay closer to the science while they design.

So go read the filing for yourself. This week’s patent is US 12531136 B2, filed by Deep EigenMatics LLC.