Trends from the Trenches: The Capability Jump of AI and Its Impact

By Bio-IT World Staff 

May 7, 2026 | A year ago, some would have said AI was not making an impact on drug development or computational biology. But Sonia Timberlake, R&D strategy consultant at Timberlake & Maclsaac Biopharma Consulting, argues that recent rapid innovation makes it hard to claim AI isn’t impacting computational biology anymore.  

“The really hopeful and exciting thing I see is how impactful AI is in our sphere,” she said in the latest episode of Trends from the Trenches. 

The biggest near-term win is not magical end-to-end drug discovery, but the unglamorous work that determines speed: bioinformatics pipelines, exploratory data analysis, figure generation, and rapid iteration across RNA-seq, single-cell, and spatial transcriptomics. 

The gap lies between what agentic coding tools can do and what scientific teams actually adopt. In mainstream software development, many companies now report the majority of code is AI-assisted. In scientific codebases in biotech and pharma, it’s the opposite. The opportunity is immediate for well-documented workflows: no one needs to hand-write a standard RNA-seq pipeline when models have abundant training examples and the real expertise lies in curating inputs and interpreting outputs. AI also expands scope, letting a small computational biology team attempt adjacent domains like basic protein modeling or data wrangling that previously required a specialist. 

Trust is the bottleneck, which is why benchmarks matter. Timberlake distinguishes benchmarks for capabilities, tasks, and processes. Capability benchmarks test narrow skills like chart and figure understanding (multimodal chart QA remains behind text). Task benchmarks test end-to-end work that resembles a real PhD-level assignment, such as analyzing spatial transcriptomics to find treatment-control differences with many nontrivial steps. Process benchmarks are the frontier. There is the question of whether AI can reliably assemble regulated, source-verified deliverables that pull from many teams, such as a pre-IND packet, without hidden errors. Open-source efforts, including FutureHouse’s BixBench and Genentech’s CompBioBench, help teams measure models under different constraints and compare “vanilla” tools against customized agents. 

Those same ideas apply when evaluating AI products sold into pharma. A practical diligence question is simple: show the benchmarks and show internal metrics that track improvement over time. If no perfect third-party benchmark exists, a serious vendor can still demonstrate objective measurement, ablation tests (what happens if you remove the proprietary layer), and baseline comparisons (determining if and why simpler models such as logistic regression failed). Without that rigor, buyers risk paying for polish instead of performance, especially because AI can generate convincing interfaces and reports that mask weak underlying methods. 

To learn more about where biotech investment is flowing, Timberlake’s workshop at this month’s Bio-IT World Conference & Expo, and more, listen to the Trends from the Trenches podcast.