制藥公司的老板們認為，開發創新性的救命藥物，需要有足夠的投資回報。但最近，藥物研發的投資回報率卻低得可憐。據德勤（Deloitte）統計，2017年，12家規模最大的生物制藥公司藥物研發部門的投資回報率只有3.2%。2010年的投資回報率還達到了10.1%。

制藥公司如何擺脫這種困境？一條途徑可能是利用人工智能在早期（失敗風險最高的階段）提高新藥發現的效率。德勤表示：“[人工智能]可用來分析來自臨床試驗、病歷、基因檔案和臨床前研究的海量數據，從中識別出模式及趨勢并提出假設，其效率遠遠超過單純依靠研究人員。”

默克（Merck）、賽諾菲（Sanofi）和阿斯利康（AstraZeneca）等大型制藥公司已經將人工智能技術引進到了實驗室當中。2017年，阿斯利康與馬薩諸塞州的初創公司BERG建立了合作伙伴關系，利用后者的人工智能平臺尋找帕金森癥等神經疾病的生物靶標和藥物。

如何利用人工智能？ BERG公司CEO尼文·R·納拉因表示，首先要“回到生物學上來”。從健康者和患者身上提取組織樣本，進行各種分子分析，結合臨床數據，然后通過BERG的人工智能平臺找出靶標。

納拉因表示，在進行數據分析時，BERG會避開“公開的數據庫。”他說道：“我們使用貝葉斯方法，而不是神經網絡。并不是把一批數據放到模型里然后得出某種相關性這么簡單。開始的時候并沒有一個預先決定的假設，而是把所有數據都輸入系統，讓數據自己生成假設。”

簡而言之，人工智能聽起來像是一門不錯的古董科學。很難想象！（財富中文網）

譯者：劉進龍/汪皓?

Creating innovative, lifesaving medicines, say pharmaceutical company bosses, requires a sufficient return on investment. But lately, that ROI stinks. In 2017, according to Deloitte, the 12 largest bio- pharma companies got a mere 3.2% return out of their drug-research arms. In 2010, that number was 10.1%.

How can pharma break out of this rut? One avenue might be the use of artificial intelligence to improve drug discovery at the earliest stages (when the risk of failure is also the highest). “[A.I.] can help analyze large data sets from sources such as clinical trials, health records, genetic profiles, and preclinical studies; within this data, it can recognize patterns and trends and develop hypotheses at a much faster rate than researchers alone,” says Deloitte.

And Big Pharma names like Merck, Sanofi, and Astra- Zeneca, are already taking it to the lab. In 2017, AstraZeneca struck a partnership with BERG, a Massachusetts startup, to use the latter’s A.I. platform to home in on promising biological targets and possible agents against neurological diseases such as Parkinson’s.

So how does it work? For starters, says BERG CEO Niven R. Narain, by going “back to biology.” Tissue samples are taken from both healthy and sick

patients, analyzed on multiple molecular levels, combined with clinical data, and then fed through BERG’s A.I. platform to suss out targets.

For analyzing that data, BERG eschews “the publicly available databases,” says Narain. “We use a Bayesian approach rather than a neural network,” he says. “It’s not just taking a bunch of data, putting it through a model, and coming up with some correlation. You don’t start out with a predetermined hypothesis—you feed the system all this data and allow the data to generate the hypotheses.”

So, in short, A.I. sounds like good old-fashioned science. Go figure.