When AI isn’t enough: How physics is shaping the next wave of drug discovery 

Over the past few years, using artificial intelligence (AI) in drug discovery has stopped being a novelty; instead, it has become part of the standard toolkit across many early-stage programs. AI-designed molecules have increasingly moved into clinical trials, and more start-ups and pharma teams use AI for target identification or virtual screening. But the game changer might be the combination of AI and physics.

A second wave is changing how the industry thinks about AI for drugs. Rather than simply applying generic AI models to databases of known molecules, some companies now look to combine computational physics, structure-based modelling, and generative AI to design novel, first-in-class molecules from scratch.  

Among them, Aqemia argues that the real dividing line isn’t between “AI-powered” and “non-AI” discovery: it’s between generic tools and approaches grounded in physics. “Over the past few years, AI in drug discovery has shifted from hype to commodity. AI has become a standard part of early discovery workflows and is no longer differentiating; the real competitive edge must lie elsewhere,” said Maximilien Levesque, Aqemia’s CEO. 

So, if AI is now the baseline, the real question becomes: where does differentiation come from? 

From history to inventing molecules: why physics matters 

Much of the early promise of AI in drug discovery rested on models trained from existing databases of known molecules, binding assays. In practice, however, this dependency on historical data presents clear limits. As a recent review on AI-driven drug discovery found, such models often struggle when data is sparse, biased, or inconsistent, a frequent situation for novel targets such as non-canonical binding pockets or proteins with little prior assay history. 

“One major misconception is that AI models trained purely on historical data can drive true drug invention — you don’t invent by remixing history.” When discovery depends only on remixing past molecules, you rarely leave the bounds of what has already been tried. 

By contrast, approaches that build in physics, atomic-scale modeling, binding energy estimation, structural dynamics, and conformational flexibility allow a different kind of exploration. Recent advances in structure-based drug design highlight precisely this shift of going beyond historical patterns: combining computational physics or knowledge-based priors with machine learning to improve the validity, novelty, and drug-like properties of candidates. 

Such physics-guided workflows make it possible to begin a discovery project even when experimental data are minimal, for example, when a protein structure is available, but no ligand history exists. They introduce information about how a molecule could bind, move, or remain stable at the atomic scale, offering guidance that doesn’t depend entirely on previously tested compounds. This enables exploration of molecules unlikely to arise through classical medicinal-chemistry derivations. 

“In my view, true invention comes from two sources: human ingenuity or physics, the first-principles science that lets us reason at the atomic scale. Without that foundation, we can’t unlock real creativity in drug discovery, nor hope to tackle targets that have never been drugged before,” said Levesque. 

Where the physics and AI combo helps 

Ask Aqemia where AI is most needed today, and the answer isn’t “more models” — it’s three enduring bottlenecks in early-stage drug discovery. 

Drug discovery today increasingly faces targets that traditional methods struggle with proteins with unusual shapes, dynamic conformations, or no precedent of bindable pockets — the “undruggable” or “difficult-to-drug” proteins. According to a 2023 Nature review of efforts to drug such challenging proteins, many of these targets, including transcription factors, protein–protein interfaces, and allosteric regulators, are now being revisited because of the advances in structural biology and computational design. 

Aqemia underscores this shift; Levesque noted that therapeutic targets today are “significantly more intricate than those addressed in the past, such as never-drugged pockets or highly dynamic proteins,” and that these require “a deep mechanistic understanding that traditional approaches struggle to capture.” 

For such targets, conventional AI trained on historical data often fails; there may simply be no historical binders to learn from. That’s where physics-informed approaches become critical. 

Also, drug discovery programs often balance the need for speed with limited experimental data. Generating high-quality assay data can be slow, costly, and time-consuming, and for novel targets, data may be very scarce. AI’s reliance on existing datasets becomes a liability when data is sparse or biased, leading to models that don’t generalize well beyond known chemistry landscapes. 

Levesque argues Aqemia’s physics-based method removes or mitigates this dependency: “This is why we use a physics-based approach, teaching first principles to an AI so it can invent molecules without mimicking existing ones.” 

Even if computational pipelines generate promising molecules, the true test remains biological validation: do they bind, modulate the target, behave in cells or in vivo, and meet developability constraints? Despite impressive computational advances, many proposals fail when confronted with real biological systems. Structure, dynamics, off-target effects, and in vivo context remain major hurdles. 

“We remain humble toward human biology. Validation remains a major bottleneck: translating computational advances into robust and reproducible biological outcomes is still the ultimate determinant of success,” said Levesque. 

Inside Aqemia’s QEMI engine: structure-first, physics-guided invention 

Aqemia begins most projects with a structure-based assessment of the target: its conformations, binding pockets, dynamics, and physicochemical constraints. This is part of a broader shift in early discovery, where companies increasingly build models around 3D information instead of starting from large ligand datasets. 

QEMI, the company’s invention engine, runs an iterative loop between generative AI and physics-based evaluation. Aqemia describes it as combining machine learning with “quantum-inspired physics” to test and refine large numbers of candidates. “QEMI explores and evaluates vast numbers of candidate molecules in silico and iterates until we converge on molecules with the right potency, selectivity, and developability, well before synthesis,” explained Levesque. 

Aqemia says the platform brings the most value when fast iteration and broad exploration are needed, where they can “crash-test many programs in parallel” to see which ones merit deeper investment and which are the best fit for QEMI. In hit finding, physics-guided scoring aims to improve the quality of initial hits rather than increase volume. Recent studies show that hybrid generative with structure-based workflows can reduce false positives and produce more drug-like scaffolds. 

In terms of range, Aqemia reports working across oncology, immunology, the central nervous system (CNS), and challenging target families, including kinases, phosphatases, methyltransferases, and transmembrane proteins. 

Internal and partnered programs, positioning as a biotech rather than a tool vendor 

Lesveque says most of Aqemia’s work is carried out through internal discovery programs, which make up the bulk of its portfolio, where the company sets the scientific agenda and selects targets it considers important or underexplored. “Internal projects allow the company to explore entirely new target classes, starting from the disease biology and working backward to the molecule.” 

Alongside this internal pipeline, the company also runs partnered programs. These collaborations are described not as software deployments but as shared scientific efforts. Levesque noted that these usually work under “shared-risk models” and that their “core technologies aren’t black-box AI but atomic-scale physics,” which leads to joint discussions on physics, biophysics, chemistry, and translational considerations. 

A notable example is the collaboration with Sanofi, announced in late 2023. The partnership centers on combining Sanofi’s therapeutic expertise with Aqemia’s physics-and-AI platform to accelerate early small-molecule design. AI-assisted discovery companies increasingly position themselves as drug-creation partners rather than software vendors.  

In that sense, Aqemia follows a trend: platforms that rely on physics or structure-based design tend to operate more like biotech companies, with pipelines, risk-sharing partnerships, and long-term research and development (R&D) commitments, rather than like pure software providers. Will this model allow AI to live up to the hype in drug discovery?