AI is enabling the development of a next generation of drugs that can more precisely target cancer cells while sparing healthy tissues.
Antibody-based therapies such as antibody-drug conjugates and T-cell engagers continue to show great potential in the treatment of solid tumors. So far, six ADCs and two TCEs are FDA-approved for solid tumors, and with dozens more in different stages of clinical development, it’s clear that innovation in these targeted therapies is really gaining momentum.
Yet despite this progress, on-target, off-tumor toxicity remains a major challenge. Ideally, all therapies would target unique markers found exclusively on tumor cells, enabling efficacious treatment with minimal collateral damage. But in reality, such ideal targets are exceedingly rare.
Most ADCs and TCEs in fact target tumor-associated antigens (TAAs). While more strongly expressed on tumor cells, tumor-associated antigens are also present on healthy cells, resulting in the targeting of both tissue types. This complicates the delicate balancing act between achieving therapeutic efficacy and minimizing the risk of unintended harm.
A familiar story is that these molecules either fail to reach clinical trials or, when they do, they show promising efficacy but are ultimately halted due to unacceptable levels of toxicity. Even when these drugs are approved, patients often have to contend with side effects.
There are ways to more precisely target tumor cells. AtLabGenius Therapeutics, where I am chief scientific officer, we focus on doing just that, developing the next generation of immunotherapies—drugs that are both safe and highly effective. Identifying the right molecules for this role is a job for machine learning.
A Better On/Off Switch
Some of the current methods for enhancing selectivity—such as pH-sensitive binding, avidity-driven selectivity and masking—have led to encouraging advances. However, to drive meaningful progress, we must aim for complete on/off killing selectivity. Solving this challenge would not only reduce the toxicity burden too often associated with antitumor therapies but would also expand the pool of viable targets to those that are also present (albeit at lower levels) on healthy cells.
To reach this goal, we need to go beyond designing molecules that simply bind to a target. Instead, we must develop antibodies that are capable of sampling the cell surface to distinguish between tumor cells and healthy cells, then eliminating the former.
To discover antibodies capable of such sophisticated interactions, researchers and industry leaders are increasingly turning to machine learning. One application of this technology, de novo antibody design, generates novel binders that target antigens and epitopes. Once binders are in hand, machine learning can aid the sequence optimization of monospecific antibodies. Although there is still some uncertainty about exactly what the companies pursuing these approaches have achieved, the growing scientific momentum of AI’s use in drug development has attracted substantial commercial investment.
At the same time, there’s a notable absence of academic and commercial efforts leveraging machine learning to engineer the design of multispecifics. Often the work of assembling a multispecific is treated as trivial, with the constituent parts stitched together using human-led, rational design.
However, the specific arrangement of parts in a multispecific is actually far more complex than it may appear. For example, we have tested TCEs containing three copies of an anti-TAA and one anti-CD3. When keeping the constituent parts in the exact same arrangement, only changing the length and rigidity of the linkers between them, we find vastly different outcomes: 100 femtomolar potency with >1-million-fold selectivity, 50 femtomolar potency with ~100-fold selectivity or no detectable potency at all. This is just one example of many that highlight how the relationship between antibody configuration and its resulting function is often non-intuitive. Therefore, there is an increased likelihood of uncovering optimal molecule designs when machine learning has a seat at the table.
Due to their modular nature, multispecifics can be built to take advantage of the difference in TAA densities between tumor cells and normal, healthy cells—a biological mechanism known as avidity-driven selectivity. In other words, the higher the TAA density, the “stickier” the behavior of our anti-tumor molecules. Existing methods typically require a large difference in TAA expression to differentiate between the tumor and normal cells. In contrast, our machine learning-driven approach enables us to work with differences between TAA expression as small as threefold. This opens up a vast new target space where the differential expression was previously not large enough to avoid tumor-targeting antibodies also killing healthy cells.
Applications beyond ADCs and TCEs
There’s no doubt that the adoption of AI is an exciting prospect for drug discovery. For us, multispecific antibodies are a particularly promising area of focus. Their modular format and programmable nature make them a good fit for AI-driven approaches. At the same time, due to their sophisticated mechanisms of action, optimizing such antibodies across multiple properties is an incredibly complex task—and brute-force screening methods simply aren’t up to the job.
Looking ahead, it’s easy to imagine this approach being applied to a range of antibody modalities beyond ADCs and TCEs, such as degrader antibody conjugates, CAR T-cells and NK-cell engagers, as well as non-tumor applications such as selectively targeting viral infections or autoimmune diseases.
Machine learning will allow us to optimize increasingly complex antibody formats, with sophisticated modes of action, ultimately raising the bar for safety and efficacy. Over the next 5–10 years, I believe we’ll begin to see these AI-designed treatments progress through clinical trials and move toward widespread adoption.
