The Future of Life Science Learned from Autonomous Driving

After years of anticipation, self-driving cars have finally arrived. While the ambitious initial goals and inflated expectations haven't been fully met, significant progress has been made.

Fully and partially autonomous vehicles now represent one of the most mature applications of artificial intelligence (AI). Today, in San Francisco, hailing a Waymo with your phone will quietly summon a white, sensor-laden Jaguar I-Pace, driverless, within minutes. Furthermore, driver-assistance features in regular vehicles, such as automated lane keeping, continue to advance rapidly.

As the autonomous vehicle industry matures, talent with interest and aptitude is steadily flowing to other AI applications, including life sciences. We believe this influx of talent can act as a catalyst for advancements in the life sciences. AI in life sciences will also likely follow a similar path of initially inflated expectations followed by an accumulation of incremental successes that redefine the industry.

Based on this, we believe four lessons from the autonomous vehicle industry can be applied to AI for drug discovery and development.

Transition to Higher-Dimensional Data Representations

As fundamental data and models improve within a given domain, more decisions are shifting from explainable, human-defined algorithms to ‘black box’ models. This is also occurring in the development of autonomous vehicle software, where the industry is increasingly adopting learned representations of the scene around the vehicle and learned models to control more functions in the stack.

For example, instead of hard-coding an algorithm to identify whether a car is parked (based on relative speed with the autonomous vehicle), the model can predict the state based on all the labeled training data of parked cars.

The autonomous vehicle industry has transitioned from development based on strictly defined requirements to modeling the vehicle’s operating domain as a complex, high-dimensional space that can be statistically represented and explored. This has led to improved performance, but decreasing interpretability.

We believe a similar shift is coming to life sciences. The complexity and diversity of biology (including complex interactions between diseases and populations) contains far more information than can be captured in human-understandable representations. To reflect this complexity in decision-making, the pharmaceutical industry will transition from simpler data representations to higher-dimensional ones.

For example, crucial questions like “Which patient population is most likely to benefit from my drug candidate?” will shift from intuitive or heuristic-based decisions to data-driven decisions. And the algorithms answering these questions will rely on complex, high-dimensional data representations.

Building an Automated Data Engine

The autonomous vehicle industry is moving away from training on limited, internally collected data toward a model trained by the vehicles themselves. Models are continuously improving based on data collected as customer vehicles drive in the real world. Data collected from customer vehicles is now one of the most valuable assets an automaker possesses in the race toward autonomy. Automated data collection through a large-scale data engine is the only way to gather the diverse data needed to build understanding of the complex domains where vehicles are deployed.

We expect the same dynamics to occur in life sciences. This ‘real-world data collection’ will occur across the discovery, development, and clinical phases:

• Large-scale in silico design and virtual screening of therapeutics
• Automated robotic wet labs integrated with AI scientists
• Scalable curation, preservation, and iterative learning from past clinical trial results
• Unique sources of rich clinical data, including multimodal, high-fidelity patient-level data
• Automated analysis and insights from market data (e.g., adverse events, patient experience with the product, medical information requests)

An automated data engine automatically improves products and processes with data gathered across all stages of drug development.

Success will require iterative and massive data collection. Each experiment and prescription will be an opportunity to collect more data for future development. That is, data gathered at each stage should be used to improve all stages of drug discovery and development.

Regulatory and safety requirements in life sciences and the heterogeneity of biological data mean that this data engine will look different from its autonomous vehicle counterpart, but the fundamental goal—deep and automated integration between data collection and product development—remains the same. Successful companies will see future iterations of their products automatically improve based on newly collected data.

Plan for the Long Term

Success in autonomous driving took longer than planned. The winners were in a position to invest heavily in development for over a decade. The same is true for AI in life sciences. The current excitement is a forward-looking opportunity for talent and capital to flow into this area. However, the winners will be those who plan for a long-term effort with little immediate payoff.

That is, a long-term effort is required. Do you have investors, partners, or customers who can support the funding needed for this journey? Can you withstand multiple clinical failures while the data engine is built as a platform?

Consider the Level of Automation

Autonomous driving, while not perfect, uses a useful taxonomy for classifying the level of autonomy. The SAE levels of driving automation classifies automated systems into six levels, including Level 0 (no automation), Level 2 (partial driving automation), and Level 5 (full driving automation).

Driver assistance and fully autonomous driving require different approaches and investments. Therefore, all autonomous driving programs have different requirements and plans depending on the level of automation being developed. The SAE classification of autonomous driving is used when expressing this concisely.

Different automation goals for different programs are true in life sciences, too. Which part of the pipeline will you automate? Some companies will automate a small part of the discovery process, such as in silico screening, and adopt traditional drug development methodologies for the rest of the stack. Others will aim to build platforms targeting fully autonomous and iterative drug discovery and development.

Different approaches will require different architectures, tools, and capital structures. Defining levels of autonomy will allow the industry to have a shared language and roadmap beyond the general goal of ‘AI adoption’.

The Road Ahead

The road ahead will be long and arduous. There will be many failures, but success will redefine the possibilities of life sciences.

Reference

https://www.convoke.bio/from-self-driving?utm_source=substack&utm_medium=email