AI Summary
→ WHAT IT COVERS Daphne Koller, founder and CEO of InSitro, provides a calibrated assessment of AI drug discovery in 2024, explaining how her company combines 28 petabytes of cellular and human cohort data to advance three clinical programs in NASH, ALS, and dry AMD, while countering claims that AI will cure all diseases within five years. → KEY INSIGHTS - **AI Data Reality:** Biological AI cannot replicate the large language model approach of scraping existing internet data. Between 50–80% of published scientific literature is non-reproducible, making raw data unreliable for model training. InSitro built a proprietary wet lab data factory generating 28 petabytes of purpose-built cellular and human cohort data, because fit-for-purpose data collection is a prerequisite before AI can generate meaningful drug discovery insights. - **Two-Pillar Discovery Platform:** InSitro's target identification relies on two anchored data sources: high-throughput cellular perturbation screens where every gene in the genome is individually disrupted to observe disease consequences, and large human cohort datasets from sources like UK Biobank and All of Us. Human genetics serves as the causal thread connecting cellular observations to whole-organism outcomes, reducing reliance on poorly translatable rodent models. - **ALS Target Strategy:** Rather than targeting TDP-43 directly — an approach that failed because clearing aggregates does not restore nuclear function — InSitro identified a buffering mechanism that allows motor neurons with TDP-43 pathology to function like healthy neurons. In lab experiments, severed neurites regrew after drug treatment in diseased neurons but not untreated ALS neurons, representing a functional recovery signal relevant to 97% of ALS patients who share TDP-43 pathology. - **NASH Program Differentiation:** InSitro's lead clinical candidate targets IRS-1 via liver-targeted GalNAc-siRNA, filing an IND this year with first patient enrollment planned for early 2027. GLP-1 drugs show only modest, transient liver fat reduction and are ineffective for lean MASH patients, who accumulate liver fat regardless of body weight. The siRNA modality offers three-to-six month dosing intervals and avoids systemic IRS-1 inhibition, which would carry significant safety risks. - **Brown Fat Quantification Breakthrough:** Using whole-body MRI analyzed through AI, InSitro measured brown adipose tissue at population scale for the first time, discovering that MRI scans taken in winter show measurably more brown fat than summer scans — confirming cold-activation hypotheses. Genetic analysis of brown fat variation then identified targets; one candidate caused mice to lose 25% fat mass without lean mass loss and without reducing food intake, indicating a caloric expenditure rather than restriction mechanism. - **Specialized vs. General AI Models:** Off-the-shelf LLMs and general computer vision models underperform on biological data modalities like microscopy, cardiac MRI, and OCT scans. InSitro uses specialized models adapted for each data type, with LLM-based agentic systems handling workflow orchestration, QC automation, and literature integration. This hybrid approach compressed experimental design cycles from six-to-nine months down to approximately eight weeks, demonstrating where general AI provides genuine productivity gains versus where domain-specific models remain essential. → NOTABLE MOMENT Koller describes how InSitro's data scientists discovered that brown fat volume in UK Biobank MRI scans varies measurably by season — winter scans show more brown fat than summer scans. This incidental finding, only visible through AI analysis at population scale, validated cold-activation biology and directly led to identifying novel cardiometabolic drug targets. 💼 SPONSORS [{"name": "AlphaSense", "url": "https://alpha-sense.com/thelongrun"}, {"name": "DashBio", "url": "https://dash.bio/pricing"}] 🏷️ AI Drug Discovery, NASH/MASH Treatment, ALS Therapeutics, Brown Fat Biology, Biotech Data Infrastructure, Clinical Stage Biotech