Where synthesis meets intelligence
AI has accelerated and validated early-stage drug discovery; continuous flow chemistry now proves critical to accelerating the scale-up of those molecules, says Subir Chakraborty of CD Chem Group
In June 2025, Insilico Medicine reported Rentosertib, a smallmolecule TNIK kinase inhibitor for idiopathic pulmonary fibrosis, that was identified end-to-end by machine learning( ML). A 71-patient Phase IIa trial published in Nature Medicine showed that the 60 mg daily dose improved lung function by approximately 98 mL compared to the placebo decline. The candidate reached clinical trial in roughly two and a half years. 1 That milestone stands as proof of concept.
As of 2026, however, no drug designed by artificial intelligence( AI) has gained regulatory approval, and around 90 % of candidates entering clinical development still fail, much the same as before. 2 AI has thus transformed early discovery economics without yet changing latestage odds, where human biology, not chemistry, determines the outcome.
Bottleneck now upstream
A working AI pipeline for anti-cancer new chemical entities( NCEs) now runs in five stages( Figure 1): 1. Target identification via‘ omics’( genomics, transcriptomics, proteomics and metabolomics) analysis
2. Generation of virtual libraries in millions of SMILES strings
3. ML-driven filtering on binding affinity and absorption, distribution, metabolism, excretion and toxicity( ADMET)
4. Synthesis and preclinical validation with results logged in electronic laboratory notebooks
5. Retraining of models on real bench data
Graph neural networks, diffusion models and reinforcement learning drive molecular design; AlphaFold 3 has transformed structure prediction. 3 Teams report reaching a preclinical candidate in less than 15 months rather than the historical time of over five years before advancing to Phases II and III.
Critically, AI filtering applies probability scoring to each candidate: predictions that a compound has, for example, an 80 % likelihood of viable oncology activity narrow millions of theoretical structures to 10 – 15 lead compounds for synthesis. This filtering step compresses the cost equation. Reducing screening burden mitigates the largest cost lever available in pharmaceutical synthesis: step count, cumulative yield loss and precious metal catalyst usage.
Synthesis still the challenge
Oncology NCEs are structurally demanding. Their cores are heterocyclic( quinazolines, pyrimidines, indoles), fluorination is widespread for metabolic stability and typical syntheses run to 10 – 20 steps. Many of these reactions are precisely those that batch processing handles least well.
Continuous flow chemistry addresses this directly( Table 1). Nitration is strongly exothermic and hazardous at scale in batch but flow reactor control removes that risk. Hydrogenation benefits from continuous hydrogen dosing and improved catalyst efficiency. Halogenation and fluorination gain precise reagent control and fewer side reactions. Cyclisation improves selectivity. Multi-step telescoping combines steps without isolating intermediates, compressing a 10 – 12- step synthesis into six to eight.
Since many oncology actives are high-potency APIs( HPAPIs) requiring containment, avoiding intermediate isolation also reduces operator exposure to cytotoxic material. Scale-up becomes time-based rather than volume-based, improving reproducibility for regulatory filing, whilst lower solvent use and reduced waste serve environmental, social and governance( ESG) objectives.
Liposomal encapsulation and next-generation delivery systems
Figure 1- Closed-loop AI discovery pipeline for anticancer NCEs
34 SPECIALITY CHEMICALS MAGAZINE ESTABLISHED 1981