Dropped connections to bolt when connecting to instances behind Network Load Balancer in AWS

We have three instance behind load balancer in AWS:

neo4j community 4.1.3
Read only - we are not writing anything to these databases, they are periodically refreshed

TCP listeners have been setup for ports 7474 and 7687

We are using spring data service with OGM to access database and it works fine with single machine direct access, but when accessing through load balancer we experience connection drops:

here is the entry from debug log:

WARN [o.n.b.t.p.HouseKeeper] Fatal error occurred when handling a client connection, remote peer unexpectedly closed connection

I have switched to http driver, and it worked fine, but bolt is faster, and I would like to make it work.
Has anyone worked with load balanced read only instances?
If yes, what parameters were used for Load Balancer (Classic vs Network AWS ELB, anything like that)