The recent uproar over the birth of genetically modified baby twins in Chinahas given the world reason to pause. Dr. Jiankui’s “pandora’s box” has led to fresh bio-ethics debates everywhere. Unfortunately, the backlash will likely result in increased scrutiny of medical research as well as the biopharma industry at large. It has happened before.
The #crisprbabies controversy is just a reminder that healthcare, whether in the form of fringe science or even legitimate medicine, presents many risks- to patients, to doctors, to investors. Everyone.
For patients, no procedure or pill is completely safe. There are everyday reminders of this. A “routine” surgery requires a signature — a recognition that death is a possible outcome. Prescribed drugs may not actually work, and they could even actually cause harm. Who among us at this point has not seen the disclaimers of side effects like death or feelings of suicide.
For doctors, the risk means higher malpractice insurance. Often this is passed on to patient. For biotech companies, risk equates to higher drug development costs, the result of rising legal fees and increasingly prolonged testing and clinical trials.
For the most vulnerable, the unlucky among us fighting pain or disease, these problems lead to sky-rocketing healthcare bills and longer delays for possible relief or even a cure.
But this seems to be the way of things in biopharma. Can we do better ?
The Future Is…Computers !
Undaunted by these challenges, the industry has looked to novel places to mitigate risk. In drug discovery, the goal is to develop safe and effective drugs, cheaper and faster.
One area of exploration has focused on high performance computing (HPC). The idea here is to do as much as possible with computing, from identifying and designing “seed” molecules to target disease, all the way to predicting outcomes of clinical trials.
Efforts in computational drug discovery started in the seventies, and that train hasn’t since stopped. There have been plenty of skeptics, and they often point to the massive investments in this area, with little results to show.
The computational drug discovery hype engine alive and kicking in 1981.
Some of these past computational forays in biopharma include:
- “Systems biology” attempted to mathematically model complex biological systems. One promise was the ability to “virtually” test drugs on simulated human physiology, thus reducing the burden of actual human trials. Unfortunately, research in this area has dwindled significantly. It has yet to produce a substantial result.
- “Molecule dynamics” attempts to model the electro-static forces between atoms and molecules to help with virtual drug design. Real synthesis of molecules is expensive and, in the end, may not produce a viable drug. So the right computational model can point researchers in the right direction. Molecule dynamics had a very slow start for several years, but recent applications in protein-protein interaction and protein folding have shown tremendous promise.
Artificial Intelligence: Version 4
The latest disruptive tech is, of course, AI, and everyone is trying to figure out how it will impact their industry, including biopharma. AI has made many promises in the past, but failed to deliver. The consequences were at times dire for the entire field of AI.
No. That’s not AI.
That said, this latest wave of AI ( the fourth one since the 50s ) is in many ways different than its predecessors. It deserves the benefit of our collective doubt. This is because today’s AI shows up at an unprecedented time in our digital history:
- Data storage is now a commodity — and we’ve been storing a lot of it. Petabytes of medical records, medical images, doctor’s notes, clinical trial data, medical research papers — all of it waiting to be mined for patterns, insights, and discovery.
- New machine learning algorithms are being developed that can learn effectively from lots of data. Deep learning, in particular, has already shown its effectiveness in medical imaging and disease prediction.
- The general availability of inexpensive computing infrastructure for running these AI workloads. All of the major public cloud providers ( Microsoft Azure, Google Cloud, Amazon AWS ) allow you to rent the most advanced computing hardware. You don’t need to work at a research or government lab to perform cutting edge AI.
- A global commitment to artificial intelligence. US can no longer boast the technical lead here — countries like China, India, Canada, and France have made AI application and research a national priority.
There is a renewed sense of exploration in biopharma, and AI finds itself at the epicenter. AI is reenergizing not just interest in computational drug discovery but also massive investment- both in big pharma and startups.
An active conversation has started to take place, and many of these conversations are happening at the intersection of technology and science.
For example, a few weeks ago, I attended the AI Applications Summit in BioPharma conference at Harvard Medical School. It was a intimate gathering of about 120 executives and researchers and practitioners in biopharma, from universities as well as industry.
Some of the highlights included:
- Harry Glorikian (author of “Moneyball For Medicine”) kicked things off with a great recap of the history of computational techniques in biopharma, as well as summarizing the current opportunities and challenges.
- Alan S Louie of IDC Insights, talked specific about ROI and how to measure the success of AI adoption in biopharma. Organizations can reap the benefits of robotic process automation (RPA) right now, including the adoption of NLP techniques.
- In addition to big pharma companies, a few startups spoke at the event. I’ve been following Recursion Pharmaceuticals for a little while now, and it was great to hear Recursion talk about their vision. I chatted with their VP Ron Alpha during a breakout session. Recursion may have one of the most sophisticated automated pipelines for capturing and analyzing cellular images.
- There were several discussions about how AI can be applied to the clinical trial process. This includes patient selection, ensuring protocol adherence, as well as techniques for processing the massive amount of medical notes that result from drug trials.
- Alex Zhavoronkov of Insilico Medicine gave an amazing talk about the use of cutting edge deep learning techniques, including Generative Adversarial Techniques for drug discovery. I’ll be talking more about this in the next blog.
This summit was a 2 day event and it was packed with excellent content. I don’t work in biotech (I work in hardware-accelerated data science ), but I was able to follow most of the presentations. The organizers have managed to get in one venue many of the leading thinkers and implementors in biopharma AI. Everyone is very approachable and there are plenty of networking and breakout events where you can have one-on-one conversations.
One of the topics discussed at the conference echoes conversations in other AI circles, such as in vehicle autonomy and automated cybersecurity. And that issue is AI trust. Two areas of concern are AI bias and AI transparency. Can we trust the decisions that an AI agent makes ? Can we trust “black box” algorithms ?
Do you trust your AI with your life ?
I think the issue of AI bias comes down data collection and data curation. Data-driven techniques like machine learning are quite good at reflecting the data that feeds it. It’s often said of bad machine learning models — “garbage in, garbage out.” Where fairness is concerned, it’s “bias in, bias out.” A stunning example of this is Amazon’s AI-based recruiting tool that demonstrated bias against female candidates.
AI doesn’t tell you why.
Now, assuming one could level the playing field in data, we still have a problem. And that’s AI’s lack of transparency. A machine learning algorithm does not tell you how it came to its decision. Why was the candidate rejected ? Why was the loan not approved ? Why is the cell not cancerous ? The model does not say. And if you go about looking under the hood, all you’ll see are millions of weights and parameters, a mess of floating point numbers.
The stakeholders in biopharma AI will need to address these issues. If they go unresolved too long, big pharma may decide that AI adds additional risk to drug discovery instead of mitigates risk.
So…What’s Next ?
In light of these issues and doubts, the innovators believe AI will become the most important computational tool in biopharma. Will AI help discover better, cheaper, and safer medicines ? Some say it’s not a question of “if,” but “when.”
But it’s not likely to bomb-drop on us — a big-bang-like, Kurzweillean “singularity” event in which algorithms suddenly output the blueprints for super drugs. Even the most optimistic AI futurists are no longer so naive. The AI revolution in biopharma will be a sequence of small wins over the course many years, with a few setbacks and lessons learned along the way.
Numerate is taming the highly combinatorial design space of small molecule drug discovery.
We are seeing some of those wins now. Google’s DeepMind just recently demonstrated an amazing result in protein folding prediction. The biopharma startups are perhaps doing some of the most innovative work. In my next blog, we’ll take a closer look at Numerate, Recursion Pharmaceutical and InSilico Medicine. They are disrupting computational drug discovery using the most advanced machine learning and data science available today.