AI Drug Discovery Innovations by Potential Sciences in 2026

Table of Contents

Potential Sciences aims to transform drug discovery process

  • Potential Sciences says it will generate massive synthetic chemistry datasets by simulating molecular interactions using physics, rather than relying mainly on lab experiments.
  • Founder and CEO Hatem Helal told WIRED the goal is to make non-viable molecules “fail fast” with better predictors of biological behavior.
  • The startup says it raised seed funding in the tens of millions of dollars from SovAI and Twin Path Ventures, among others, according to WIRED.
  • Researchers quoted by WIRED say accurate simulation could be transformative—but protein engineering remains extremely hard.

What is Potential Sciences and its focus in drug discovery?

Potential Sciences is a British startup betting that artificial intelligence can “speedrun” drug design by shifting more of the early discovery workload from wet labs to computation, according to WIRED. The company launched on November 5, WIRED reported, with a plan to train AI models on large-scale chemical datasets it generates itself.

The core focus is prediction: building systems that can forecast how candidate drugs will behave as they interact with human biology.
Potential Sciences’ approach is positioned as a response to a bottleneck in the field: many AI drug discovery efforts build on the same foundational public dataset of protein structures, the Protein Data Bank. WIRED describes the Protein Data Bank as a public database of 3D atomic structures that researchers have expanded over decades with results from new experiments. Helal argues that this shared base is too small to train models that can reliably accelerate the search for viable candidates, according to WIRED.

If the company’s bet works, Helal told WIRED, cheaper experimentation could make it economical to pursue drugs for rare conditions that have previously gone untreated—and could support “bespoke treatments” tailored to an individual’s biology.

How does Potential Sciences plan to build large-scale chemical datasets?

Potential Sciences plans to create synthetic data by simulating chemistry from the ground up—starting with interactions between atoms, then small molecules, then more complex amino acids and proteins—using basic principles of physics, according to WIRED. The idea is to generate far larger training datasets than those available from public experimental repositories alone.

Helal’s argument, as quoted by WIRED, is that drug discovery is fundamentally a search-and-prune problem across an enormous space of possible molecules. He illustrated the scale with a cosmic comparison: “Of all the atoms in the universe, there’s around 10^60 ways that you could assemble them into molecules that you could potentially give to a human,” he told WIRED. In that view, better tools are those that navigate faster and discard more of the space earlier.

The company’s dataset strategy is also a critique of the status quo. WIRED reports that a large proportion of AI drug discovery labs develop models on top of the Protein Data Bank’s 3D atomic structures. Potential Sciences is trying to break that dependency by manufacturing a much larger pool of training examples through simulation.

Academic experts interviewed by WIRED see the conceptual appeal. Alan Fersht, emeritus professor of chemistry at the University of Cambridge, told WIRED that if scientists could accurately calculate interactions between molecules in the cell—and the fluctuations in their structures—then predicting binding would “indeed revolutionize drug development.” He added: “I am optimistic that this will eventually happen,” according to WIRED.

What funding has Potential Sciences received and from whom?

Potential Sciences says it has seed funding “in the tens of millions of dollars” from SovAI—the UK government’s AI venture capital fund—and Twin Path Ventures, among other investors, according to WIRED. WIRED also reports the founders declined to disclose the precise figure or valuation.

SovAI’s involvement matters not only as capital but as industrial policy. The UK government established SovAI earlier this year with a mandate to invest in British startups that could become “indispensable components of the global AI supply chain,” providing the country geopolitical leverage, according to WIRED. The fund has roughly $800 million at its disposal and can offer portfolio companies access to the UK’s national AI supercomputer, WIRED reported.

Joséphine Kant, head of ventures at SovAI, told WIRED that Potential Sciences’ physics-first approach differentiates it: “Coming at drug discovery from a physics angle—the physics of chemistry—sets [Potential Sciences] apart from other companies.” Kant joined the startup’s board as part of the investment, according to WIRED. She also emphasized the practical advantage of compute access, telling WIRED that with those resources the company can work “at a scale that would be very difficult otherwise for a small startup.”

WIRED also noted that by receiving investment from SovAI, Potential Sciences joins other AI drug discovery companies Basecamp Research and Isomorphic Labs, as well as materials discovery startup CuspAI.

How does the approach of Potential Sciences aim to reduce drug discovery costs?

Potential Sciences’ cost thesis is straightforward: replace a portion of costly, time-consuming lab experimentation with simulation-driven prediction, then use AI to prioritize only the most promising candidates for real-world testing, according to WIRED. Helal told WIRED the company is betting it can train models capable of identifying highly targeted drugs for specific diseases while bypassing large amounts of early experimentation.

The mechanism is “fail fast,” but in a specific sense. Helal told WIRED the company wants molecules that won’t work in the clinic to fail early, enabled by “better predictors of how they will behave as they interact with human biology.” In drug discovery economics, earlier failure is often cheaper failure: fewer assays, fewer animal studies, fewer cycles of redesign.

Helal also tied lower experimentation costs to neglected markets. If it becomes cheaper to explore and validate candidates, he told WIRED, pharmaceutical companies could find it economical to pursue drugs for rare conditions that have previously gone untreated.

At the industry level, there’s a tension between AI’s promise and the reality of overall R&D spending. Deloitte’s 2026 study, summarized by CompaniesHistory, reported that the average cost to develop a single asset rose from $2.23 billion in 2024 to $2.67 billion in 2025. That same CompaniesHistory summary also notes that industry leaders acknowledge AI has improved early-stage success rates but has not yet delivered broad reductions in total R&D costs.

Potential Sciences is effectively arguing that better predictive accuracy—especially around binding and biological behavior—could shift that equation, at least in the earliest stages, as described by WIRED.

What is the significance of identifying non-viable drug candidates quickly?

The significance is less about speed for its own sake and more about where drug development burns money and time. Helal’s “fail fast” line to WIRED captures a common industry pain point: late discovery-stage surprises can cascade into expensive downstream work.

In Potential Sciences’ framing, the critical capability is prediction—knowing earlier whether a molecule is likely to behave well when it meets human biology. “That means having better predictors of how they will behave as they interact with human biology,” Helal told WIRED. If those predictors are reliable, teams can prune the candidate set aggressively and focus lab resources on fewer, higher-quality hypotheses.

But WIRED also highlights a hard boundary: even if AI improves candidate selection, it does not erase the cost and time of proving safety and efficacy. Peter Wolynes, a professor of chemistry at Rice University, told WIRED that designing a drug is “nothing compared to figuring out what it will do to mice; to really sick patients; to normal people.” He added that, in some sense, design is “not the rate-limiting phenomenon,” according to WIRED.

That distinction matters when interpreting “fail fast.” Faster elimination in silico can reduce wasted lab cycles, but it doesn’t automatically compress clinical timelines. It can, however, reduce the number of weak candidates that ever reach expensive testing—an outcome aligned with Helal’s stated goal in WIRED.

What limitations exist in current AI drug discovery efforts?

One limitation is data scarcity—at least for the kind of model training many teams want to do. WIRED reports that a large proportion of AI drug discovery labs build on the Protein Data Bank, a public database of 3D atomic structures expanded over decades through experiments. Helal argues that, despite its importance, the dataset is still too small to train AI models advanced enough to reliably accelerate the search for viable drug candidates, according to WIRED.

Another limitation is the sheer difficulty of the underlying science. The central technical question for Potential Sciences, WIRED notes, is whether it can predict how well drug candidates will bind to disease-related proteins as accurately as it hopes—a “marked leap forward” would be required. Wolynes offered a blunt assessment of the field’s maturity: “the engineering is at the level of the Romans, not of the Renaissance, in terms of how protein engineering is done,” he told WIRED.

There are also practical and commercial constraints. WIRED reports that Potential Sciences has not settled on a revenue model for the technologies it is developing, nor decided whether to make its models freely available.

Finally, there is a dual-use concern. Helal told WIRED that tools designed to minimize toxicity could be flipped to maximize it: “If we can design tools that perfectly minimize toxicity, it’s a bit flip to maximize it as well.” That risk complicates decisions about openness, access, and deployment.

How does Potential Sciences intend to improve AI model training for drug discovery?

Potential Sciences’ training strategy is to feed AI models far larger datasets than those typically available from public experimental sources—by generating synthetic data through physics-based simulation, according to WIRED. The company’s bet is that scale and coverage of chemical interactions will translate into better predictive performance, especially for tasks like forecasting binding and biological behavior.

The logic is rooted in the mismatch between the size of chemical possibility space and the size of existing labeled datasets. Helal’s quote to WIRED about ~10^60 possible assemblies underscores why he sees drug discovery as a navigation problem: models need enough examples to learn how to prune effectively.

SovAI’s Kant connected this ambition to compute. She told WIRED that giving Potential Sciences access to computing resources enables work “at a scale that would be very difficult otherwise for a small startup.” That matters because simulation-based dataset generation is computationally expensive by design: the company is trying to compute interactions rather than measure them in a lab.

Experts interviewed by WIRED suggest the upside is real if accuracy is achieved. Wolynes, while skeptical about the maturity of protein engineering, still told WIRED of Potential Sciences: “I’m sure it will be an improvement.”

The company’s approach also sits within a broader wave of AI optimism in science. WIRED notes that Demis Hassabis of Google DeepMind has argued that advances in AI will eventually lead to momentous scientific breakthroughs, and that many startups hope drug discovery will be among the first frontiers transformed.

The Future of AI in Drug Discovery: A Comprehensive Overview

Understanding the Role of AI in Modern Drug Development

AI’s role in drug discovery is increasingly described as foundational rather than experimental. Drug Target Review argued in 2026 that AI had become a core driver of discovery workflows, with computational prediction used alongside experimental validation to test hypotheses earlier and accelerate candidate selection.

Yet the economics remain contested. In that context, Potential Sciences’ pitch—reduce experimentation costs by improving prediction, as described by WIRED—targets a specific slice of the cost curve rather than the entire end-to-end process.

The near-term reality, as Wolynes told WIRED, is that safety and efficacy testing remains a dominant time and cost driver. The long-term hope, echoed by Fersht in WIRED, is that sufficiently accurate simulation could change what “early certainty” looks like in molecular design.

Key Players and Innovations in AI Drug Discovery

Potential Sciences is entering a crowded market, and WIRED notes it will face steep competition from entrepreneurs and academics attacking similar bottlenecks from different angles. SovAI’s portfolio context also signals how governments are trying to shape strategic advantage: WIRED reported that SovAI has invested in AI drug discovery companies Basecamp Research and Isomorphic Labs, and in materials discovery startup CuspAI.

Outside this specific startup, the broader sector shows both momentum and selectivity. CompaniesHistory summarized a 2024 analysis of approximately 300 AI-native biotechs, reporting 67 AI-discovered molecules among 6,147 drugs in clinical development and Phase I success rates of 80–90% for AI-discovered molecules—figures it presented as above historic industry averages.

The open question for Potential Sciences, as WIRED frames it, is whether simulation-generated data and physics-first modeling can deliver the “marked leap forward” in binding prediction accuracy that would justify the ambition—and whether that improvement translates into real-world therapeutic outcomes, not just better models.

Related Posts

Scroll to Top