No PriorsNo Priors Ep. 140 | With Benchling Co-Founder and CEO Sajith Wickramasekara
CHAPTERS
- 0:00 – 2:08
Benchling’s origin story and what the company does today
Sajith Wickramasekara explains how his experience moving from software into a biology lab revealed how outdated lab workflows were. He outlines Benchling’s mission as “modern software for scientific progress,” and describes its scale across biotech, pharma, and academia.
- •Motivation: replacing paper notebooks/spreadsheets with modern collaborative tooling
- •Benchling supports design, execution, organization, analysis, and sharing of lab work
- •Company scale: ~1,300 biotech/pharma companies and ~7,000 academic institutions
- •Benchling’s vantage point across large pharma and cutting-edge AI biotechs
- 2:08 – 3:49
What data Benchling manages across the drug R&D lifecycle
Sarah asks what kinds of data actually live inside Benchling beyond sequences and assay outputs. Saji frames drug development as thousands of interdependent steps, with heterogeneous data that must remain searchable and connected across programs.
- •Drug development spans target selection, molecule design/optimization, preclinical tests, clinical trials, manufacturing, and regulatory work
- •Benchling focuses on “lab-originated” scientific data across modalities and stages
- •Captures relationships between molecules, experiments, and results (including scale-up/process data)
- •Core value: organize and make data searchable to support better scientific decisions
- 3:49 – 6:37
Biotech macro cycle: post-COVID exuberance to a “dot-com bust” trough
They zoom out to the biotech market cycle and why the past few years have been especially painful. Saji attributes the downturn to a confluence of capital costs, geopolitics, regulatory uncertainty, and over-exuberance around new therapeutic modalities.
- •Biotech cycles are real; recent period likened to the dot-com bust
- •COVID/mRNA peak drew in generalist capital and platform-company enthusiasm
- •Rate changes, regulatory uncertainty, tariffs, and geopolitics contributed
- •Hype vs reality gap for modalities like gene editing, cell/gene therapies, and RNA
- 6:37 – 9:30
China’s rise in biotech: speed, cost, and pharma’s shift to buy molecules there
Saji argues the next decade will be defined by speed and cost, where China is increasingly competitive. They discuss big pharma licensing or buying Chinese-origin molecules, including a concrete example of a successful therapy now used in the US.
- •China’s advantage: faster/cheaper early clinical development, even in newer modalities
- •Top pharma increasingly sources molecules from China vs US biotechs
- •Example: J&J partnership with Legend Biotech; Carvykti as a recognizable success
- •Western reactions range from competitiveness to skepticism about standards/replication
- 9:30 – 13:08
Why drugs feel ‘magical’—and why they’re still so slow, expensive, and failure-prone
Saji summarizes his essay’s core premise: drugs are high-ROI innovations that become cheaper over time, yet the pipeline remains brutally inefficient. He emphasizes that late-stage failures and artisanal processes drive massive costs and long timelines.
- •Prescription drugs are a relatively small share of US healthcare spend yet deliver huge value
- •Drug economics: generics drive long-run cost declines unlike many healthcare services
- •Typical reality: ~10 years and ~$2B+ to market; many programs fail late
- •Biotech remains ‘artisanal’ digitally: bespoke workflows, fragile institutional knowledge
- 13:08 – 16:13
AI’s real opportunity: improve each step and compress feedback loops (not just ‘type disease → get drug’)
They discuss where AI can meaningfully reduce cycle time and improve decision quality. Saji argues clinical trials aren’t the only lever; better molecules and faster learning loops matter most, and AI should be applied across the many steps after discovery too.
- •Clinical trials are costly, but many failures trace back to ‘not good’ molecules
- •Goal: compress years-long feedback cycles to get learning faster
- •AI value is stepwise: improve many bottlenecks, not a single end-to-end magic button
- •Manufacturing/process development becomes an even bigger constraint if discovery speeds up
- 16:13 – 18:35
Benchling AI explained: simulations in workflow + agents for deep research and automation
Saji lays out Benchling AI’s two pillars: integrating predictive models directly into lab workflows, and deploying agents that automate research and reporting over structured Benchling data. A case study illustrates how agents can recover lost institutional knowledge and avoid redundant experiments.
- •Simulation layer: bring open-source/proprietary/internal models into scientists’ day-to-day workflows
- •Aim for ‘right model at the right moment’ usable by wet-lab scientists without heavy compute skills
- •Agent layer: deep research agent that reasons over Benchling’s data model/context
- •Example: found prior mouse-model work buried in old notebooks, saving ~months of effort
- 18:35 – 21:13
AI scientist vs copilot: near-term augmentation, long-term autonomy
They debate the ‘AI scientist’ narrative and what will realistically happen in the next 1–2 years. Saji expects an augmentation trajectory (copilots) before full autonomy, with trust, accountability, and workflow integration as limiting factors.
- •‘AI scientist’ evokes end-to-end autonomous design-make-test-analyze loops
- •Near-term bet: augmentation like copilot; autonomy needs time, money, and patience
- •Analogy: Waymo (full autonomy) vs Tesla (incremental capability gains)
- •Accountability and risk management keep humans in the loop, especially in high-stakes domains
- 21:13 – 22:52
Adoption barriers in life sciences: trust, accuracy, security/IP, and the missing ‘chat’ interface
Saji notes that although model capabilities are advancing quickly, most R&D orgs aren’t heavily using AI yet. Concerns about accuracy, IP, security, and compliance loom large, and the industry still lacks a breakthrough interface that makes AI truly frictionless in scientific workflows.
- •Life sciences have ‘GPT but no chat’: capabilities exist, usability/integration lags
- •Primary blockers: accuracy, IP leakage, security, legal/compliance concerns
- •Adoption varies by geography/culture; concerns increase outside SF
- •Winning AI will be the one embedded in workflows and trusted by users
- 22:52 – 23:58
What pharma is doing now: pilots, pragmatic transformation, and data advantage for model training
Sarah asks how pharma views AI today. Saji describes cautious optimism: many pilots, limited org-wide transformation so far, but a major advantage in pharma’s ability to generate large-scale experimental data to train models.
- •Pharma has excitement and executive attention (e.g., ‘AI officers’) but moves methodically
- •Widespread pilots (Copilot-like tools) without full R&D transformation yet
- •Pharma’s differentiator: large-scale experimental data generation for training proprietary models
- •Expectation: more unique predictive models emerging from pharma compute teams
- 23:58 – 27:05
Which models are useful today—and why ‘AI-discovered drugs’ is the wrong yardstick
They cover the current ecosystem of bio models and the emergence of open source, plus new collaboration structures like federated learning. Saji pushes back on the simplistic critique that AI hasn’t produced a marketed drug, arguing impact should be measured by how many experiments and decisions AI improves across the pipeline.
- •Open source in biology has become durable; new structure/biology models are proliferating
- •Examples mentioned: structure prediction and developability-focused models; federated approaches (e.g., Lilly TuneLab)
- •Big gap remains: models that predict patient outcomes and clinical success reliably
- •Better metric: growing share of experiments ‘touched’ by simulation/prediction/AI
- 27:05 – 29:28
Business models and data markets: tools vs assets, model distribution, and transacting on trustworthy data
They explore whether AI will change biotech business models and how model companies may evolve. Saji highlights a future of scalable model distribution (more SaaS-like), and increased data transactions enabled by normalization and trust, including the possibility of sharing negative data.
- •Model-only businesses may struggle as model-building commoditizes; some may become biopharma with pipelines
- •Alternative: distribute models broadly via platforms with pay-per-use / SaaS-like economics
- •Biotech oddly lacks data transactions despite data being its ‘currency’—trust and formatting are blockers
- •Tooling/normalization could unlock markets for preclinical and even negative data sharing
- 29:28 – 34:52
Why digital tools matter in pharma: the slow shift from paper to structured data—and AI as a tailwind
Sarah challenges the belief that only drug assets create value. Saji argues physical tools companies have proven the tooling thesis, while digital lagged due to on-prem/paper inertia; Benchling’s long push was to get science online and structured—now AI makes the payoff obvious.
- •Physical tooling giants (e.g., instrumentation/services) show tools can create massive value
- •Life sciences lagged cloud adoption; early Benchling years were ‘bring science online’ evangelism
- •Next decade: convincing orgs to adopt structured data as a system of record
- •AI makes benefits more immediate, accelerating willingness to modernize workflows
- 34:52 – 42:53
Company building in a regulated vertical: betting on AI, staying close to customers, and aligning incentives
They shift to leadership and organizational choices, including a major internal bet to focus a co-founder on AI. Saji emphasizes customer intimacy as his most reliable operating ‘algorithm’ and discusses the challenges of blending academic science culture with commercial software incentives.
- •Strategic bet: reassigning leadership focus to AI despite market downturn fears
- •Founder ‘moral authority’ helps drive controversial pivots and absorb reputational risk
- •Core operating advice: talk to customers continually; PMF moves as markets shift
- •Interdisciplinary management: reconcile academic incentives (papers, cheap labor) with commercial reality (sell software, make money)
- 42:53 – 48:13
Tech ↔ biotech cultural exchange, plus personal AI excitement beyond Benchling
Saji closes with what each industry can learn from the other: biotech should communicate more directly and humanize impact, while tech should absorb biotech’s rigor around safety and validity. He ends with personal enthusiasm for agentic coding tools and AI accessibility for non-technical users, including family and education.
- •Biotech/pharma should ‘go direct’ and tell patient/scientist stories to improve public understanding
- •Tech can learn rigor, validation, and safety constraints needed for regulated, high-stakes work
- •Personal excitement: agentic coding brings back the joy of building quickly
- •AI’s natural-language UX expands access (parents, tutoring/learning, reduced ‘IT support’ burden)
