Hot Topics in Chemistry 2025: New Directions in the DMTA Process

While at the conference, I also made sure to catch as many presentations and poster sessions as I could. The ACS meeting is a great opportunity to learn about new developments and stimulate thinking about how to address challenges.

A persistent question in drug discovery has been how can the process be improved. In 2012, Eroom’s Law was first coined, the observation that drug discovery is becoming slower and less productive in contrast to the accelerating pace of microprocessor technology described by Moore’s Law.

Scientists have formalized the chemical discovery process into the iterative Design-Make-Test-Analyze cycle. With traditional tools and approaches, that process typically takes many years to identify a drug candidate. I was eager to learn about innovations in the chemical discovery process to cut down that timeline.

One could consider two approaches, and both received attention at the ACS. Below are some observations and key themes that stuck with me from the meeting.

The bottleneck step in the DMTA cycle is usually making compounds, and many groups are using automation to solve these problems.

From my past experience, good tools exist for automating most of the necessary unit operations, but integration is a challenge. Tying together reaction setup, execution, isolation, and purification is impressive.

Working at the right scale to do this efficiently was a theme throughout the Novartis and JNJ/Janssen session. They showcased their parallel automated synthesis systems and examples of success with their platform. They targeted 1-10 mg of final compound, which is plenty of material to gain information and drive projects forward at the hit-to-lead phase.

From my past experience, good tools exist for automating most of the necessary unit operations, but integration is a challenge. Tying together reaction setup, execution, isolation, and purification is impressive.

Working at the right scale to do this efficiently was a theme throughout the Novartis and JNJ/Janssen session. They showcased their parallel automated synthesis systems and examples of success with their platform. They targeted 1-10 mg of final compound, which is plenty of material to gain information and drive projects forward at the hit-to-lead phase.

Among the challenges in synthesis efficiently is finding suitable reaction conditions. Automation of reaction setup in parallel is simple with liquid handlers, but analysis by LCMS is a serial process limited by run times.

The Blair group at St. Jude developed a direct mass spectrometry method (avoiding chromatography) to determine reaction success/failure by observing diagnostic fragmentation patterns. Throughput is approximately 1.2 s / sample (compared to >1 min / sample by LCMS). This allows a 384 well plate of reaction mixtures to be analyzed in 8 min.

Another tact is to use existing data to predict whether new reactions will work.

Connor Coley and his group at MIT are building models to predict reaction success, and in comparison to expert chemists, the computational models provide similar accuracy. As part of a compound design workflow, this could be used to quickly weed out problematic target compounds and replace them with similar structures with greater likelihood of success.

Time constraints prevented me from attending the session on Direct-to-Biology, but isolation and purification of compounds is usually a bottleneck between the Make and Test phases. A robust connection between parallel synthesis to assay biology could be a game changer.

The best way to perform fewer iterations is to design better compounds. The prevailing hope is that AI steps in here.

While academic groups are performing this necessary basic research, generative AI is already taking hold in industry. Shanthi Nagarajan presented Eli Lilly’s system, which is designed to output structures with good activity at the target, drug-like properties, novelty, and synthetic feasibility.

In Lilly’s comparison with more traditional enumeration + ML-scoring approaches, their generative models output a set of candidate compounds that exclusively met their definition of “drug-like”. Earlier workflows gave only ~1% compounds with the desired properties, and 99% of compounds were filtered out.

For someone trained as a synthetic chemist, the point about synthetic feasibility was a welcome advance.

The last time I went to a conference with discussion about AI for drug discovery, compounds presented from generative AI systems ranged from close analogs of existing drugs to structures with questionable or unsynthesizable features. Through the talks at ACS, presenters either explicitly discussed synthesizability or showcased reasonable looking structures. The focus on bringing theory to practice bodes well for realizing the promise of AI in drug discovery.

One prediction of AI’s impact is to reduce the time to identify a clinical candidate from 6 years to 1 year. With such potential, this year’s ACS devoted 1.5 days to advances in generative AI for chemical discovery (including beyond the drug space).

The field is still exploring what types of models are most suitable under different circumstances, in addition to what representations of compounds are appropriate for generative AI.

The national meetings of the American Chemical Society are always equal parts exciting and overwhelming, and the Spring 2025 edition in San Diego was no different. I left with a pile of papers to read, new faces to follow in the industry, and enthusiasm for the next event.